AI is reshaping software, services, and knowledge work at extraordinary speed.

But there is an important limit to what AI alone can do.

No matter how intelligent an AI model becomes, it still runs on classical computing infrastructure. It can optimize, approximate, compress, and reason over problems, but it remains bounded by the computational constraints of classical architectures.

There are entire categories of economically important problems where the bottleneck is not intelligence, but computation itself:

• molecular simulation

• catalyst discovery

• materials science

• cryptography

• large-scale optimization

• physical system modeling

• advanced semiconductor design

These problems sit underneath trillion-dollar industries:

• pharmaceuticals

• energy

• manufacturing

• logistics

• defense

• semiconductors

This is where quantum computing matters.

Not because it replaces AI.

Because it expands the frontier of computable problems that AI and classical systems can eventually operate on top of.

The strategic importance of quantum computing has therefore always been clear. The challenge has been that the industry has remained stuck between scientific promise and commercially viable infrastructure.

That is why Atom Computing’s newly published paper may end up being remembered as one of the most important quantum computing milestones in years.

The real bottleneck in quantum computing

Most people still evaluate quantum computing companies through the wrong lens.

The conversation tends to revolve around:

• qubit counts

• fidelity percentages

• benchmark demonstrations

Those metrics matter, but they are no longer the central challenge.

The real bottleneck is whether a quantum system can operate continuously while managing:

• error correction

• qubit degradation

• qubit loss

• system orchestration

• operational scaling

In other words:

can a quantum computer function like infrastructure rather than a laboratory experiment?

This is much harder than it sounds.

Quantum systems are extraordinarily fragile. Operations introduce errors. Measurements disturb the system. Components degrade while computation is happening. The qubits literally disappear during computation.

This is where most architectures begin to break down.

To become commercially viable, quantum computing platforms ultimately need to demonstrate several things simultaneously:

1. the ability to continuously detect and fix errors

2. systems that can keep running reliably during long computations

3. protection against failures and instability

4. the ability to scale to much larger systems over time

5. continuous operation without constant resets or interruptions

6. resilience when individual components fail or degrade

7. flexible system designs that can adapt as the technology evolves

Very few systems have demonstrated even subsets of these capabilities together.

Atom Computing just demonstrated many of them simultaneously.

The Quantum Computing Breakthrough the Industry Has Been Waiting For — Five Major Advances

Today the Atom Computing team announced what I believe is one of the most important milestones the quantum computing industry has seen in years: the first complete demonstration of continuously operating quantum error correction in a neutral atom system.

With this result, Atom becomes one of two companies to demonstrate many rounds of performant quantum error correction, and the first neutral atom company to do so. The only other company to publicly demonstrate comparable sub-threshold error correction behavior to date has been Google using superconducting qubits.

Importantly, Atom demonstrated something even more significant than isolated logical operations.

The company showed that as the system became larger and more redundant, logical error rates improved rather than degraded.

In plain English:

the system became more reliable as it scaled.

Historically, this has been one of the defining bottlenecks in quantum computing. Larger systems often introduced so much additional noise, coordination overhead, and instability that performance worsened instead of improving.

Atom demonstrated the opposite while simultaneously showing continuous qubit replenishment, sustained operation, arbitrary qubit connectivity, and logical memory persisting beyond the lifetime of the underlying physical qubits themselves.

That combination is what makes this paper such an important breakthrough for the broader industry.

To double click into the major advancements demonstrated in the paper:

1. Repeated Quantum Error Correction During Live Computation

Atom demonstrated repeated error correction cycles using a toric code architecture with true mid-circuit measurement, reset, reuse, and replacement of qubits.

This is foundational to fault-tolerant quantum computing.

Useful quantum systems cannot rely on perfect physical qubits. They require logical qubits continuously stabilized through repeated cycles of error correction while computation is actively running.

Historically, many systems could demonstrate isolated logical operations or short-duration memory experiments. Atom demonstrated repeated correction cycles integrated into a continuously operating system architecture.

In plain English:

the system was actively detecting and correcting problems while computation continued.

That is one of the core requirements for building commercially useful quantum systems.

2. A Sub-Threshold Quantum Error Correction Result

Atom demonstrated what the industry refers to as “sub-threshold” behavior.

The company showed that increasing the size and redundancy of the error correction system reduced logical error rates across multiple operating regimes.

This is one of the most important milestones in fault-tolerant quantum computing because historically, adding more qubits and more correction overhead often introduced so much additional complexity and noise that systems became worse, not better.

Atom demonstrated the opposite.

As the system became larger and more redundant, logical performance improved.

In plain English:

the system improved as it scaled.

Google’s 2025 surface code result became one of the most important quantum computing breakthroughs in recent years because it demonstrated this behavior in superconducting qubits.

Atom has now demonstrated similar sub-threshold behavior while simultaneously solving broader systems-level problems around continuous operation and replenishment.

Importantly, Atom achieved this using a neutral atom architecture with arbitrary connectivity rather than the fixed planar architectures common in superconducting systems.

One of the most important charts in the paper. Historically, larger quantum systems often became less reliable as complexity increased. Atom demonstrated the opposite: larger, more redundant systems achieved lower logical error rates.

3. Continuous Mid-Circuit Reloading and Qubit Replenishment

This may be the most important architectural breakthrough in the paper.

Neutral atom systems face a uniquely difficult challenge:

• atoms heat during operation

• atoms are lost during measurement

• atoms physically disappear from the computational array

Instead of treating this as a fatal limitation, Atom designed the architecture around continuous replenishment.

The system continuously:

• detects missing or degraded qubits

• replaces them

• reloads fresh atoms from an external reservoir

• preserves logical operation while computation continues

The paper demonstrates that without replenishment, the system eventually degrades and logical operation collapses.

With replenishment, the system stabilizes into steady-state operation over many cycles.

Imagine if servers inside a data center could fail, be physically replaced, and computation could continue uninterrupted without restarting the system.

That is conceptually what Atom demonstrated at the qubit level.

This is one of the clearest demonstrations yet of a quantum architecture beginning to behave like continuously operating computational infrastructure rather than a fragile laboratory experiment.

The most important chart in my opinion in the paper. Without atom replenishment, the system degrades and eventually fails. With replenishment, Atom demonstrates steady-state operation. This is one of the first demonstrations of a quantum architecture behaving more like infrastructure than a laboratory experiment.

4. Arbitrary Qubit Connectivity Through Dynamic Atom Movement

Most quantum architectures remain constrained by local connectivity. Qubits can only directly interact with nearby neighbors, creating substantial overhead and coordination complexity as systems scale.

Atom demonstrated an architecture built around dynamic movement of neutral atoms, enabling arbitrary qubit connectivity across the computational system.

A useful analogy is transportation infrastructure.

Most quantum architectures today behave like cities with rigid road systems where information must move block by block through neighboring intersections.

Atom’s architecture behaves more like an air traffic network where qubits can dynamically move and directly interact where needed.

As systems scale, that flexibility matters enormously.

Flexible connectivity can materially improve:

• quantum error correction efficiency

• logical gate implementation

• operational overhead

• scalability of larger code families

The paper explicitly highlights that this architecture may enable practical utility-scale codes inaccessible to many nearest-neighbor planar systems.

This is one of the most important long-term advantages of neutral atom architectures and one of the least appreciated by the broader commercial community.

5. Logical Memory That Outlived the Physical Qubits Themselves

Finally, Atom demonstrated logical memories that outlived the underlying physical qubits composing the system.

Physical atom lifetimes were roughly 10–15 seconds.

Logical information persisted:

• 44 seconds at distance-3

• 225 seconds at distance-7

That result is conceptually profound.

The computation survived longer than the physical hardware components themselves.

In effect:

the computation survived the death and replacement of the underlying hardware.

That is one of the defining characteristics of true fault-tolerant computing systems.

It means the architecture is beginning to preserve computation independently of the survival of the underlying physical qubits.

The logical memory survives dramatically longer than the underlying atoms themselves. The computation effectively outlives the hardware components composing it, one of the defining characteristics of fault-tolerant computing.

Why this is such a significant breakthrough

The importance of this paper is not that Atom demonstrated one isolated benchmark.

It is that the company integrated multiple foundational capabilities into a single continuously operating architecture:

• repeated error correction

• sub-threshold behavior

• mid-circuit measurement

• qubit replacement

• continuous replenishment

• arbitrary connectivity

• sustained logical memory

Historically, most quantum computing breakthroughs have remained relatively narrow:

• improved qubit fidelity

• isolated logical operations

• shallow memory demonstrations

• benchmark circuits

Atom demonstrated something much closer to operational infrastructure.

And importantly, the company did so using a neutral atom architecture that possesses several long-term advantages:

• no cryogenic infrastructure

• highly scalable qubit counts

• flexible connectivity

• modularity

• dynamic qubit movement

This is the first time a neutral atom architecture has started to resemble operational computing infrastructure rather than an isolated physics demonstration.

The broader takeaway

Utility-scale quantum computing is still ahead of us. Logical error rates remain too high for many commercially important workloads, and substantial scaling challenges remain.

But this paper materially changes the conversation around what architectures may realistically scale.

The frontier in quantum computing is no longer simply:

“Who has the best qubits?”

The frontier is increasingly:

“Who can build continuously operating fault-tolerant computational infrastructure?”

That is a systems engineering problem as much as a physics problem.

And Atom Computing just delivered one of the clearest demonstrations yet that this transition may finally be underway.

They’ve come a long way since we led their Series Seed, and the fun is just starting.

Subscribe now

That sounds contradictory, but it is the entire point.

On raw model capability, you could argue Opus now sits behind OpenAI and Google. GPT-5.5 appears to have pushed OpenAI back to the coding frontier, and Google continues to look extremely strong at the model layer.

But in practice, many developers still reach for Claude Code first.

Why?

Because coding is no longer just a model problem. It is a product problem.

The Model Is Not the Agent

A model is the intelligence layer. It predicts, reasons, writes, explains, and generates code. It is the thing you benchmark. It is the thing people compare on leaderboards.

A coding agent is different.

A coding agent is the application layer wrapped around the model. It decides how to use the model inside a real engineering workflow. It figures out what files to inspect, what context to retrieve, which commands to run, how to interpret errors, when to edit code, when to test, when to retry, and how to keep moving toward the goal.

The model answers the prompt. The agent does the work.

That distinction is easy to underestimate. A great model inside a weak product can feel surprisingly average. A slightly weaker model inside a great agent can feel dramatically better, because the agent turns raw intelligence into useful output.

This is the Claude Code effect.

Claude Code Gives the Model Hands

The base model does not inherently know your repo. It does not automatically inspect your file tree, search across your codebase, run tests, read stack traces, apply patches, manage context, or decide when to ask for permission.

That all lives in the agent layer.

Claude Code gives Opus a working environment: file access, code search, editing, terminal execution, test loops, error inspection, planning, task memory, context management, permissioning, and a developer-friendly interface.

None of that is “inside” the model in the pure sense.

The model may be smart, but without the agent layer, it is trapped in a chat box. Claude Code gives the model hands. It lets the model operate inside the environment where software actually gets built.

That can make the experience feel 10x better, even if the underlying model is not 10x smarter.

This is why a weaker model with a better harness can outperform a stronger model with a weaker product surface. The bottleneck shifts from intelligence to execution.

The Model Race Still Matters

This does not mean the model race is irrelevant. It still matters a lot.

GPT-5.5 appears to put OpenAI back at the frontier for coding. It looks especially strong at narrow, complex reasoning tasks: debugging, PR review, documentation, and working through dense code or data structures.

Claude remains highly competitive, especially for open-ended work: planning, scaffolding, greenfield development, and turning vague intent into usable software.

So this is not a simple “OpenAI wins” or “Anthropic wins” story. Different models are starting to feel better for different parts of the engineering workflow.

But the more important shift is that the model is no longer the whole product.

The Product Layer Now Matters More

A coding agent is not just a model with a text box. It is a full system around the model: the CLI, the editor integration, the context window, the tool calls, the retry behavior, the speed, the prompt caching, the permissions model, and the workflow design.

This “harness” increasingly determines whether a model feels magical or frustrating in practice.

The best raw model can still lose if the product surface is slow, clunky, expensive, or hard to fit into real engineering workflows. The inverse is also true: a model that is not obviously the strongest on paper can win usage if the system around it is better.

That is why comparing raw models is becoming less useful.

The better comparison is not just GPT-5.5 vs Claude Opus. It is Codex vs Claude Code vs Cursor vs OpenCode.

In other words: not engine vs engine, but car vs car.

Users do not buy engines. They buy cars that get them where they need to go.

The Economics Are Changing

The old question was: how much does this model cost per token?

The better question now is: how much does it cost to get the task done correctly?

That difference matters.

GPT-5.5 is not cheap on a per-token basis at $5 per million input tokens and $30 per million output tokens. But price per token is only one part of the story.

A more expensive model can still be cheaper overall if it solves the task faster, needs fewer retries, writes better code, and requires less human cleanup.

That is what companies will care about. Not token price in isolation. Task completion cost.

This is where the agent layer becomes financially important. The agent decides how much context to pull in, how many tool calls to make, how often to retry, how much output to generate, and how much work the human has to clean up afterward.

The economics of coding agents are not just model economics. They are workflow economics.

Small Technical Changes Can Have Big Cost Impact

This is where the technical details start to matter.

Tokenization can quietly change the economics. If the same codebase, prompt, or agent workflow is counted as 35% more tokens, then the effective cost goes up 35%, even if the published price per token stays the same. In coding agents, where most of the cost comes from repeatedly reading and managing context, that can materially change cost per task.

A tokenizer change can act like a hidden price increase.

Speed is becoming part of the product too. Some systems are now explicitly pricing for faster output, priority access, and latency guarantees. Reasoning depth is another lever. Models increasingly expose settings like low, medium, high, and extra-high reasoning, which means users are not just choosing a model. They are choosing a cost, quality, and latency profile for each task.

The interface is no longer just prompt → answer.

It is becoming task → reasoning budget → context strategy → tool use → output.

That is a very different product category.

Benchmarks Are Useful, But Incomplete

Benchmarks still matter. They give us a way to compare progress. But they are not the same as real engineering work.

Take SWE-bench. According to SemiAnalysis, it started with roughly 93,000 merged pull requests and was filtered down to 2,294 tasks where tests could validate the fix.

That is useful, but it also tells you what the benchmark is measuring: tasks that can be cleanly validated by tests.

Real software work is messier. Requirements are ambiguous. Context is scattered across files, tickets, docs, Slack threads, tribal knowledge, and half-remembered architectural decisions. The “right” answer is often not a single patch, but a sequence of decisions across planning, implementation, review, and iteration.

A leaderboard can tell you something.

It cannot tell you everything.

The Real Benchmark Is the Workflow

The most interesting detail in SemiAnalysis’ recent post is not just that one model beats another. It is that different systems can have very different economics and outcomes.

Prompt caching, tool-call strategy, context gathering, and input/output ratios all change the real cost of using an agent. Early traces showing Codex at roughly an 80:1 input/output ratio versus Claude Code at 100:1.

That kind of detail matters because the harness determines how much context gets pulled in, how many tokens get burned, how many tools get called, and how often the human has to intervene.

In practice, that may matter more than a few points on a benchmark.

The real benchmark is not just whether the model can solve a task in isolation. It is whether the full system can move through the messy loop of software development: understand the intent, inspect the code, propose a plan, make the change, run the test, interpret the failure, fix the issue, and leave the developer with something they trust.

The Bigger Shift

The coding-agent market is moving from model performance to system performance.

The winners will combine strong models, fast inference, reliable tool use, excellent context handling, low-friction workflows, and sensible task economics.

The best model will still matter.

But the best coding product will matter more.

Coding agents are no longer just intelligence in a box. They are becoming full-stack engineering tools.

And that is why Claude Code can make a third-best model feel like a first-place product.

Most Investors Underwrite the Wrong Risk in Frontier Series A

The hardest part of a frontier Series A is not that the technology is too early.

It is that a growing share of Series A capital is still using SaaS-style product and traction mental models to underwrite companies.

But for many frontier companies, the real bottleneck is deployment risk: manufacturing, reliability, qualification, integration, procurement, service burden, and the ability to turn technical credibility into real customer commitments.

That mismatch is where a lot of good companies get stranded.

Seed can fund scientific possibility. Later capital can fund visible deployment. Series A has to fund the messy conversion in between, where the product is technically real but not yet trusted enough for a customer to commit or for the next capital pool to underwrite scale.

That is the real Valley of Death.

And it matters more now because AI has accelerated the formation of frontier companies across compute, defense, energy, industrial systems, and infrastructure. More teams are reaching technical credibility faster. But integration, qualification, procurement, and deployment still move on physical-world timelines. The result is a widening gap between what founders need to prove and what many Series A investors know how to underwrite.

What has to be true

A frontier Series A gets interesting when four things are true.

First, the technology is real enough. The product has moved beyond scientific possibility. The remaining work is mostly engineering, reliability, and deployment.

Second, customer commitment is real enough. You can point to the milestone that makes a real buyer willing to qualify, pilot, design in, pre-purchase, or commit budget.

Third, the economics are real enough. The model gets stronger with scale, payback is legible, and the company can become a strong business rather than an expensive project.

Fourth, the team is credible enough for the next gap. Not just to invent the product, but to get it across the next bottleneck.

If one of those is weak, the round usually gets hard.

Frontier is not just deep tech

A quick distinction matters here.

Deep tech usually describes the underlying technical difficulty: hard science, long R&D cycles, hardware, novel engineering.

Frontier is broader and more useful for founders. It includes deep tech, but also the harder problem that comes after invention: deployment, integration, procurement, capital formation, and where value accrues when a new system collides with the real world.

That is why this post is about frontier Series A, not just deep tech fundraising.

The hard part is usually not just building the thing. It is getting it close enough to customer reality that the market can trust it.

Where value accrues

When the bottleneck shifts from invention to deployment, value does not automatically accrue to the company with the most elegant science.

You can see it in defense. The valuable company is not just the one with a clever missile concept. It is the one that finds a corner of the trade space others missed, proves the math through independent technical assessment, hits the required price point, and becomes the only credible option in a class of system. That is when a technical program starts to look like a category of one.

You can see it in semiconductor. The breakthrough is not “more memory” in the abstract. The value accrues if a company can turn a real technical breakthrough into a manufacturable, integrated product that actually survives packaging, yield, thermal, and supply-chain reality within the timeframe that meets the customer where they need to be.

That is the actual separation.

Not between good science and bad science. Between technologies that are impressive and companies that can become unavoidable.

What founders think they are saying, and what investors hear instead

A lot of frontier Series A rounds break because the founder thinks they are telling a story about progress, while the investor is listening for proof.

The reality is, most companies are doing too much. Early stage companies do not die from starvation. They die from indigestion: too much complexity, too much capex, too much integration pain, and too little room for error.

The founder blueprint before you raise

Before you start a Series A process, you should be able to answer these questions in plain English.

Not in deck language. In underwriting language.

If your answers are vague, the round is probably too early. If your milestones do not map to customer action, it is still a seed story.

The technical questions that actually matter

As you double click, the Series A is not just about proving the product works. It is about proving the product still works once reality gets involved.

That means founders need strong answers to questions like:

• Does the system still work outside the demo?

• What degrades first at scale?

• How much tuning or human intervention does it actually require?

• What part of the advantage disappears once you include integration, ruggedization, certification, or deployment burden?

• Can it be manufactured repeatably, not just demonstrated once?

• How sensitive is performance to real-world variation?

• What technical assumption is doing most of the work in the story?

• Is this now an engineering problem, not an open science problem?

Founders often underweight this because the demo is compelling. Investors often overweight it because this is where many frontier systems break.

Customer commitment is the real milestone

One of the most common errors in frontier fundraising is using vague customer enthusiasm as proof of readiness.

Customer interest is cheap. Customer commitment is expensive.

Meetings are not commitment. Pilots are not always commitment. What matters is whether the next milestone makes a real customer action rational.

A strong Series A should be able to say, with specificity, what customer behavior changes once the next threshold is crossed.

That might mean:

• a paid pilot with defined success criteria

• a qualification program with named milestones

• a design-in or development agreement

• a pre-purchase, reservation, or conditional volume commitment

• a budgeted deployment path tied to clear specs and delivery terms

The right question is not “do customers like this?”

It is “what will a real buyer actually do once we clear the next trust threshold?”

A lot of founders try to bridge this gap with LOIs. That can be useful, but only when the LOI contains real commercial and operational substance. Most do not.

If you say you have $500 million of signed LOIs, that sounds impressive. In many cases, it means very little. A non-binding one-page LOI with no money attached, no target specs, no milestones, and no pricing is usually just a low-cost expression of interest in a future capability.

Fallacy of LOIs

If your LOIs do not cost the buyer something, they usually mean: if you someday build the thing, call us.

That is not customer commitment. It is low-cost future interest.

The key test is simple: does the customer proof make commitment more real, or just interest more legible?

Why unit economics matter earlier than founders think

A lot of frontier founders treat unit economics as a later-stage question. That is a mistake.

At Series A, investors do not need fully mature economics. But they do need evidence that the economics improve in a way that creates real operating leverage. Proof there is a business here.

If the model scales linearly or sub-linearly, has unclear payback periods, requires high capex, or leaves too little room for error, the company can still be technically real and commercially weak. If you modeled it on perfect assumptions, the entire model comes into question.

That is especially true in frontier markets because reality is almost always more expensive and slower than the model.

Founders need to be able to answer questions like:

• Which costs actually come down with scale, and why?

• Does capex unlock a structural advantage, or just more revenue with the same fragility?

• How sensitive is payback to delay, utilization, cost overruns, and deployment friction?

• What has to be true for the next capital layer to underwrite the asset?

• Do early deployments already show positive or near-positive contribution economics?

The key question is not just whether the product works. It is whether each deployment makes the business stronger.

What we look for at the frontier Series A

At this stage, we are not just asking whether the technology is impressive. We are asking whether this can become the trusted system sitting at a real bottleneck in an important market.

If a founder cannot answer these points clearly, the Series A story is probably still too early or too soft.

Closing

The real challenge in a frontier Series A is not simply proving the technology is impressive.

It is proving that the next dollars turn technical credibility into trusted deployment, real customer action, durable economics, and a company that can actually capture the value it creates.

That is the gap many investors still underwrite poorly.

And that is the gap we care most about.

If you are building in AI infrastructure, defense, energy, compute, or industrial systems, and your Series A is really about crossing that gap, we would love to meet you.

The paper quickly spread across social media, venture circles, and policy discussions. The radar chart showing the gap between theoretical capability and current usage has become a kind of shorthand for “AI is coming for your job.”

But the reaction says more about how people interpret technology than about what the paper actually shows.

The report has three fundamental problems. It is weaker than the discourse around it suggests, it frames the question incorrectly, and the way the findings are being interpreted has fueled unnecessary hysteria.

None of this means the authors did poor work. It means the conclusions people are drawing from the work are misguided.

And more importantly, it means the debate about AI and labor is happening on the wrong axis.

First: The Data Is Far Less Dramatic Than the Narrative

The first issue is simple: the paper does not show meaningful labor disruption.

The authors themselves say this directly. Looking at unemployment data across occupations with the highest levels of AI exposure, they find no statistically meaningful change in employment outcomes since the release of ChatGPT.

The one signal people frequently cite is a slowdown in hiring among workers aged 22–25 entering highly exposed professions. Even there, the paper is careful to note the effect is “just barely statistically significant.”

In other words, the evidence of labor disruption is weak.

That should not be surprising. Labor markets are slow-moving systems. Organizations do not instantly restructure around new technology. They adjust hiring, workflows, and capital investment over time.

If anything, the paper is a reminder that the labor market impact of AI has barely begun to show up in the data.

Yet the public reaction has been to treat the findings as proof of imminent mass displacement.

That interpretation simply isn’t supported by the evidence.

Second: The Radar Chart Is Persuasive but Misleading

The second issue is the way the report visualizes its core finding.

The widely shared radar chart compares theoretical AI capability with actual usage across occupations. In many fields, the share of tasks AI could theoretically perform is dramatically larger than the share currently being automated.

Visually, this creates a striking image: large swaths of economic activity appear to be sitting on the edge of rapid automation.

But that interpretation misunderstands what the chart actually represents.

The gap between capability and deployment does not mean automation is about to sweep through those sectors. It simply means organizations have not yet integrated the technology into their workflows.

That gap is not evidence of imminent labor displacement. It is evidence that we are still early in the adoption cycle.

Most general-purpose technologies show exactly this pattern. Capability arrives first. Integration follows slowly as institutions, software systems, and professional norms adapt.

The chart is not a forecast of job loss. It is a map of where deployment has not yet happened.

Third: The Real Problem Is the Framing

The deeper issue is that the report frames the problem in a way that does not match how labor markets actually work.

The analysis treats occupations as collections of tasks. It asks what percentage of those tasks AI could theoretically perform.

That may be a useful technical question, but it is not the right economic one.

Jobs are not priced by tasks.

They are priced by time or outcomes.

This distinction matters enormously.

If a job sells hours or throughput, automation reduces the value of the worker. But if a job sells judgment, responsibility, or results, automation can make the worker more powerful rather than less.

Once you look at the labor market through that lens, the impact of AI becomes much easier to understand.

Where AI Replaces Labor

Some types of work are fundamentally priced by time or volume of output.

Administrative operations, customer support, routine analysis, QA testing, market research, and many forms of entry-level knowledge work fall into this category.

In these roles, firms are essentially purchasing throughput. If the same work can be produced faster or cheaper with AI, the economic incentive is straightforward: automate.

This is where AI will compress labor demand.

But even here the transition is unlikely to appear as dramatic layoffs. More often it shows up as slower hiring and shrinking entry-level pipelines.

Companies stop expanding teams that perform routine cognitive work. Over time the workforce contracts through attrition rather than mass displacement.

Routine knowledge work becomes commoditized.

Where AI Becomes a Superpower

At the other end of the spectrum are professions where value is tied to outcomes rather than time.

Executives allocating capital. Engineers designing systems. Doctors diagnosing patients. Lawyers crafting legal strategies. Investors evaluating opportunities. Creatives determining taste and direction.

These jobs are not paid by the hour of work required to reach a conclusion. They are paid for the quality of the decision.

In these contexts, AI does not replace the worker. It increases their leverage.

The value of the work comes from judgment, synthesis, and accountability. AI simply expands the information and analytical capacity available to the decision-maker.

The best practitioners in these fields will become dramatically more productive.

Historically, technological progress tends to increase the returns to top performers. AI will likely accelerate that dynamic.

The Compression of the Professional Services Pyramid

Between these two extremes lies a large category of professional knowledge work. A post I’ve previously written about in The Great Legacy Extinction: AI's $20T Takeover of Professional Services.

Lawyers, consultants, analysts, and engineers combine judgment with a substantial amount of routine analytical work.

AI will likely reshape these professions in a different way.

Much of the work historically performed by junior employees—research, drafting, documentation, preliminary analysis—can now be done by machines.

The result is not the disappearance of the profession. It is the compression of the pyramid.

Firms will need fewer junior workers while senior professionals operate with greater leverage.

Over time, this may fundamentally change how expertise is developed. The traditional apprenticeship pipeline that trains future experts may shrink or evolve.

The Jobs AI Barely Touches

Finally, there are large parts of the labor market where AI has limited direct impact.

Work that depends on physical presence—construction, repair, agriculture, transportation, hospitality—remains constrained by the real world.

AI may improve planning, logistics, and coordination in these sectors, but the worker remains essential.

Ironically, many of the jobs economists once expected automation to eliminate may prove to be among the most resilient.

The Real Opportunity

Seen through this lens, the Anthropic report becomes much more interesting.

The most important insight is not which jobs AI can perform tasks in. It is how large the gap still is between AI capability and real economic deployment.

Across many professions, models could theoretically perform a majority of tasks, yet actual usage remains far smaller.

Closing that gap will require far more than better models.It will require new software, new workflows, new institutional norms, and new organizational structures.

In other words, the next wave of AI companies will not simply build smarter models. They will redesign how work itself gets done.

The Real Divide

The debate about AI and jobs is currently focused on the wrong question.

The question is not how many tasks AI can perform.

The question is how different types of work are economically structured.

Some jobs sell time.Some jobs sell outcomes.

AI compresses the first category and superpowers the second.

Understanding that distinction will matter far more than any radar chart.

We moved from CPU-bound systems to GPU acceleration. We optimized arithmetic throughput. We counted FLOPs. We built larger clusters and assumed the bottleneck would remain math.

That era is ending.

AI systems are shifting from compute-bound to memory-bound. The limiting factor is no longer how many operations a chip can perform. It is how quickly data can move to and from those compute units — and how much energy it costs to do so.

Recent SemiAnalysis work makes this explicit. Even when raw silicon is comparable, rack-based architecture alone can drive up to ~9x differences in effective bandwidth depending on topology and coupling. Performance is increasingly determined by how well we can move the data, not just compute it.

Nvidia CEO Jensen Huang put it bluntly: “Without the HBM memory, there is no AI supercomputer.”

This is not a chip problem.

It is a memory problem.

The Science of HBM — and Its Limits

High Bandwidth Memory exists because traditional DRAM could not keep pace with accelerators.

DRAM is a high-density, cost-efficient memory technology that stores data in capacitors requiring periodic refresh, making it ideal for large-capacity systems like HBM in AI GPUs.

SRAM is a low-latency, flip-flop-based memory technology that does not require refresh, making it significantly faster but less dense and more expensive than DRAM, and therefore best suited for on-chip caches and compute-adjacent buffers.

By stacking DRAM dies vertically with through-silicon vias and dramatically widening the data bus, HBM increases aggregate bandwidth while reducing energy per bit moved. It trades frequency for width and distance for proximity.

Without HBM, modern AI systems would stall.

But HBM is an evolutionary solution to a structural constraint.

As GPUs scale from tens to hundreds of billions of parameters — and from single-node systems to rack-scale integration — pressure on bandwidth-per-watt compounds. Compute accelerates. Data movement increasingly dominates system-level energy.

The memory wall has not disappeared.

It has been deferred.

The Constraint Is Compounding

Three bottlenecks are converging.

• Bandwidth density. Even with HBM, large models saturate memory channels. More stacks mean more packaging complexity and tighter yield constraints.

• Energy per bit. At scale, moving data often consumes more energy than computing on it. In dense racks, this becomes a thermal limit.

• System integration. Performance deltas now come from topology, interconnect, and memory coupling — not just transistor counts.

SemiAnalysis shows that recent generation-to-generation systems have delivered 10x+ improvements in performance-per-dollar for inference in certain configurations. Those gains accelerate deployment. They do not relieve memory pressure. They amplify it.

Transformer sizes have grown 240x in 2 years, while HBM 2x in year years. Compute scaling increases memory stress.

The faster we push AI, the harder the wall pushes back.

Why This Is Hard

Reinventing memory is difficult because physics pushes back. Higher bandwidth degrades signal integrity, stacked architectures amplify thermal density, and energy per bit rises with distance — all while packaging complexity erodes yield.

Memory must optimize density, bandwidth, latency, energy, manufacturability, and cost simultaneously. Improvements in one dimension often degrade another.

This is why most attempts at “better DRAM” fail.

The bar is not novelty.

It is physics plus yield plus ecosystem compatibility.

Memory Scarcity Not By Accident.

This is no longer theoretical.

Driven by AI demand, 64GB DDR5 pricing has surged from roughly $150 to $500 in under two months. Industry discussions already point to 25–30% DRAM contract price increases into late 2025. HBM capacity remains constrained and concentrated.

“Buy more GPUs” is no longer the only scaling constraint.

Memory pricing and supply are becoming strategic variables. Three suppliers — Samsung, SK Hynix, and Micron — control roughly 95% of global DRAM production. As AI demand explodes, they are rationally reallocating wafer capacity toward ultra-profitable HBM, which consumes more advanced process steps and significantly more packaging complexity per gigabyte than commodity DDR.

The incentive is clear: prioritize scarce capacity toward the highest-margin product, keep supply tight elsewhere, and let pricing follow.

In that world, memory scarcity is not an accident.

It is an equilibrium.

Nvidia’s HBM Hedge

In that context, Nvidia’s $20B acquisition of Groq makes sense. Groq’s thesis is not faster arithmetic. It is reduced dependence on off-chip memory. By keeping more of the working set in on-chip SRAM and designing the compiler around locality, Groq reduces reliance on bandwidth-hungry external DRAM and HBM.

SRAM is limited and expensive. The advantage only materializes when the entire stack — hardware and compiler — is co-designed.

That is the point.

Nvidia effectively secured an inference path that is less HBM-hungry — a hedge against a future where memory supply and pricing determine who can scale inference.

Bending the Memory Curve

Beyond attempts at reducing HBM reliance, several serious approaches are attempting to bend the memory curve:

1. HBM4 and smarter base layers: The next version of HBM uses much wider connections and adds logic at the bottom of the memory stack, turning memory into something more programmable instead of just passive storage. They extend the current architecture, but they are still fundamentally constrained by stacking thermals, packaging yield, and DRAM scaling limits.

2. Better 3D stacking techniques: New bonding methods let memory chips stack more tightly and connect more efficiently, increasing bandwidth without driving power and heat out of control. Tighter stacking improves density, but heat removal, mechanical stress, and manufacturing yield become exponentially harder at scale.

3. Doing small bits of compute inside memory: Instead of constantly moving data back and forth, some designs perform limited operations directly inside the memory stack to cut unnecessary traffic. It reduces data movement for specific workloads, but general-purpose AI models are difficult to map cleanly onto limited in-memory compute blocks.

4. Splitting fast and slow memory more intelligently: Keep only the hottest data in ultra-fast HBM and push everything else to larger, slower pools so expensive bandwidth isn’t wasted. Tiering improves utilization, but it does not eliminate the need for massive high-bandwidth memory at the core of the system.

5. Replacing some copper links with optics: As electrical wiring becomes too power-hungry at high speeds, optical connections can move data using less energy over longer distances. Optics helps at scale-out boundaries, but it does not solve the dense, short-distance bandwidth and power constraints inside a package.

6. Fixing the packaging bottleneck: A lot of memory limits now come from how chips are physically connected and cooled, not from the memory cells themselves. Advanced substrates unlock routing density, but they remain complex, yield-sensitive, and dependent on concentrated manufacturing capacity.

7. Designing software around memory limits: Compilers and system software can schedule work and compress data in ways that reduce how often chips need to reach off-chip memory. Smarter compilers reduce traffic, but they cannot overcome the fundamental physical cost of moving bits off-chip.

Every one of these approaches attacks the same constraint: energy and bandwidth at scale. They optimize around the constraint, but they do not yet eliminate it.

Building the Next Great Memory Company

The next great HBM company will not win by being marginally better.

It will win by shifting system-level throughput.

Hyperscalers do not buy memory chips. They buy rack-level performance. If a new architecture does not materially improve tokens-per-watt at cluster scale, it will be absorbed by incumbent roadmaps.

To matter, a next-generation HBM architecture must:

• Deliver step-function improvements in bandwidth-per-watt.

• Improve rack-level throughput, not isolated memory benchmarks.

• Integrate cleanly into existing accelerator and packaging ecosystems.

• Scale under realistic yield and thermal constraints.

• Offer roadmap depth that compounds advantage over generations.

Anything less will be incremental.

The Frontier Is Shifting

For the first decade of modern AI, progress was gated by algorithms.

Today, it is gated by data movement and energy.

When rack topology can drive ~9x differences in effective bandwidth, and performance-per-dollar can jump 10x generation-over-generation, it becomes clear that system-level integration — not raw silicon — is the competitive frontier.

Memory now sits at the center of that frontier.

HBM once reset the limits of accelerator design. The next breakthrough will do the same.

If you are building a next-generation memory architecture that materially shifts bandwidth-per-watt — or investing behind one — we should talk.

The frontier is no longer just software. It is physics.

And the next great AI infrastructure company will be built there.

It has always been a capability—the systems that shape the world.

• Railroads stitched a continent together.

• Electricity rewired how work and life functioned.

• Energy unlocked modern industry.

• Aviation collapsed distance.

• Semiconductors compressed computation into silicon.

• Software turned information into leverage at planetary scale.

Each frontier followed the same pattern: a technological inflection unlocked new productive capacity, reshaped the economy, and reinforced America’s ability to build, defend, and lead.

Energy, materials, manufacturing, logistics, and defense together account for roughly 80% of global GDP. They are the physical economy—the capabilities on which modern life, security, and prosperity rest.

Over the last thirty years, venture capital rationally followed what was easiest to scale. Software abstracted complexity, margins expanded without factories, and returns compounded without touching the physical world. Empires were built on screens while the harder work of building factories, infrastructure, and industrial systems quietly fell out of favor.

The consequence was not just reduced industrial capacity or slower defense readiness. It was more fundamental: erosion of technological superiority itself—and with it, rising risk to Western democracy.

Because democracy is not defended by ideals, it is defended by asymmetric capability. The next thirty years cannot look like the last.

The Last Thirty Years

The last thirty years produced extraordinary software wealth — and simultaneous industrial drift.

1. America’s industrial capacity has been overtaken. Across materials, manufacturing, logistics, energy, and defense, productivity stalled while competitors invested in integrated supply chains, modern factories, and learning at scale. In many categories, the gap is generational.

2. The United States is deeply dependent on adversarial supply chains: rare earths, specialty chemicals, advanced components, and manufacturing infrastructure. These are not peripheral goods; they are the foundations of industrial and defense capability.

3. The defense industrial base cannot keep pace with modern conflict. While adversaries iterate on software timelines, U.S. systems still take a decade or more to field. Production capacity thinned. Learning slowed. Readiness eroded.

These failures compound. Industrial weakness drives dependency. Dependency constrains defense. Slow iteration widens the gap between what technology enables and what institutions can deliver.

That gap is the American Frontier.

The Seven American Frontiers

Breakthroughs in models, sensing, robotics, and manufacturing have collapsed the cost of acting in the physical world. Geopolitics made industrial competence existential. Capital efficiency crossed the threshold where industrial systems can produce venture-scale outcomes.

For the first time in a generation, venture returns, national security imperatives, and domestic industrial policy are aligned.

Maintaining technological superiority is essential to preserving Western democracy. Investing in the American Frontier is how we intend to continue our commitment to America’s prosperity. To create meaningful work, strengthen communities, and prove that American workers can build the most sophisticated products on Earth.

The seven frontiers below, taken together, span the inputs, production systems, energy sources, coordination layers, and scaling mechanisms of the physical economy—and the majority of both U.S. and global GDP.

They are ordered deliberately—from the most fundamental capabilities to the mechanisms by which capacity ultimately scales.

1. Reclaiming Control of Critical Materials

Control over critical materials is control over whether anything can be built at all. Advanced manufacturing, energy systems, and modern defense depend on a narrow set of inputs— high-performance magnets that drive electric motors and precision guidance, specialty chemicals that enable semiconductor fabrication and battery cells, and advanced alloys that survive heat, stress, and radiation in aircraft, rockets, and reactors. These materials sit upstream of every factory, grid, and weapons system.

Today, the United States does not control them.

China processes roughly 90% of the world’s high-performance rare-earth magnets—the kind used in EV drivetrains, wind turbines, precision-guided munitions, and aircraft actuators. It dominates the refining of lithium, cobalt, and graphite needed for batteries, supplies the majority of specialty chemicals required for chip manufacturing, and controls key alloy and metallurgical steps required for aerospace and defense production. When these foundations are offshore, industrial ambition becomes theoretical. Scale slows. Costs rise. Resilience disappears.

The consequence is not just supply risk—it is design constraint. Products are engineered around what can be sourced, not what is optimal. Capacity planning becomes geopolitical guesswork. Defense readiness becomes a function of foreign processing capacity.

What has changed is not demand, but feasibility. New chemistries, software-defined processes, and AI-driven control systems are making domestic production economically viable—often with higher yields, tighter specifications, and faster iteration than legacy global supply chains. At the same time, circularity is becoming an industrial input, not an environmental afterthought. Materials embedded in scrap, e-waste, and end-of-life systems can now be recovered, refined, and redeployed with precision—turning waste streams into strategic assets.

Critical materials are not rare because they are exotic—they are critical because without them, nothing else works.

What unlocks venture-scale returns:

Proprietary process or catalyst IP that compounds with data, delivers verified yield improvements, creates non-linear scale efficiencies, supports high margins, and anchors long-term, stable offtake agreements.

What to avoid:

Linear commodity extraction or recycling businesses without process control, learning curves, or pricing power.

Emerging leaders:

Solugen, Hertha Metals, Vulcan Elements, Molg, Hgen, Redwood Materials, Boston Metal, Ascend Elements, 6K, Electra

2. Reshoring the American Factory

Reshoring manufacturing is not a geography problem—it is a design problem. Most factories today are incremental evolutions of decades-old layouts, built around manual coordination and brittle assumptions about precision. Layering software or automation onto these systems rarely changes the economics; it just adds complexity to workflows that were never designed to adapt.

Modern factories lose 30–50% of productive time to retooling, scheduling, and coordination.

The factory of the future starts from a clean sheet. Robotics, AI, and sensors are the foundation, not retrofits. Machines understand objects and physics, not just fixed paths. Systems are instrumented end-to-end, with AI coordinating machines and operators, flagging issues early, and reconfiguring workflows as inputs, yields, or capacity needs change.

Humans move up the stack—overseeing systems, handling edge cases, and training models—while machines run repetitive execution. This shifts capex and unit economics. Factories become smaller, more specialized, and closer to demand, with modular deployments instead of monolithic plants. Networks of high-throughput facilities replace single, fragile mega-factories.

Built this way, cost curves bend through iteration, not labor arbitrage. Output per worker doubles. Scrap and rework fall materially. Utilization rises as systems learn instead of resetting. Gross margins expand from commodity levels into the 30%+ range, driven by uptime and yield rather than wage compression. Manufacturing becomes resilient—able to scale commercially and support defense production without fragile labor dependency.

What unlocks venture-scale returns:

Factories designed as robotics and AI-native systems where learning compounds with every part produced, driving durable cost and throughput advantages.

What to avoid:

Incremental automation layered onto legacy factory layouts that preserve manual coordination and reset learning with each change.

Emerging leaders:

Hadrian, Freeform, Bright Machines, Seurat Technologies

3. Energy Abundance Through Decentralized Power

The legacy grid was not designed for the future we are entering. Electrification, advanced manufacturing, AI infrastructure, and defense systems are driving energy demand that centralized generation, aging transmission, and multi-year interconnection queues simply cannot meet. When power is scarce, slow, or unreliable, factories stall, utilization collapses, and strategic readiness erodes.

Interconnection queues exceed 2,000 GW in the U.S.—more than 2× peak national demand.

The next era of energy will be defined by decentralized, locally deployed power—systems that convert abundant resources directly into reliable capacity at the point of use. This includes geothermal, nuclear, solar, and battery storage, but also emerging approaches that rethink how energy is generated, transmitted, and delivered, wherever they meaningfully compress deployment timelines or expand access.

What matters is not the source, but the effect: power that bypasses grid bottlenecks, deploys on industrial timelines, and gives operators control over uptime, cost, and resilience. When energy becomes abundant and local, entirely new industrial use cases become viable—processes that were previously uneconomic, facilities that can be sited closer to demand, and production systems that can run continuously rather than defensively.

Energy abundance is not an endpoint. It is a force multiplier. It expands what can be built, where it can be built, and how fast capacity can scale. The regions that can deploy power fastest will be the regions that can actually produce, manufacture, and defend at scale.

This is not a climate thesis. It is an industrial one.

What unlocks venture-scale returns:
Decentralized energy platforms that integrate generation, deployment, and financing to deliver repeatable, reliable power independent of legacy grid constraints.

What to avoid:
Energy technologies sold as standalone hardware or bespoke projects, without ownership of deployment, uptime, or economics.

Emerging leaders:
Base Power, Radiant Energy, Antares, Standard Nuclear, Fervo Energy, X Energy

4. Rapidly Deployable Defense at Scale

Modern conflict rewards speed, adaptability, and learning—not bespoke perfection. Yet much of the U.S. defense industrial base is optimized for the opposite: exquisite hardware, decade-long development cycles, and cost structures that make iteration slow and brittle. When systems take years to field and are too expensive to replace, learning stalls and readiness erodes.

The failure mode is simple: 12–15 years from program start to deployment in a world where adversaries iterate in months.

Ukraine made this visible. Advantage did not come from singular breakthrough platforms, but from the ability to deploy systems quickly, absorb losses, adapt in weeks, and continuously upgrade capability in the field. Software-defined systems paired with rapid manufacturing allowed forces to sense, decide, and respond faster than traditional acquisition models can support.

The next generation of defense capability is not about flooding the battlefield with commodity hardware. It is about platforms designed to improve through use. AI enables adaptive behavior, faster targeting, and real-time learning across fleets. Software turns hardware into an evolving capability rather than a static asset.

The winners design for cadence: build, deploy, learn, refine. Systems must be cheap enough to learn from, fast enough to update, and intelligent enough to matter.

When defense platforms are software-defined and rapidly manufacturable, the economics change. Production depth increases, learning accelerates, and capability compounds. The defense industrial base begins to look like what it must become: a modern industrial base—competitive, adaptive, and able to keep pace with peer adversaries.

What unlocks venture-scale returns:

Defense platforms that can be manufactured quickly and cheaply, deployed in months, iterated continuously, and scaled across programs and customers as enduring systems of record.

What to avoid:

Commodity hardware plays—whether drones or other platforms—where differentiation erodes quickly, margins collapse, and advantage depends on scale alone rather than capability.

Emerging leaders:

Astranis*, Long Wall*, Skyryse*, Anduril, Adyton*, Shield AI, Epirus

5. Physical Intelligence for the Real World

Intelligence in the physical world is not a modeling problem—it is an execution problem. Unlike language or vision, the physical world is variable and unforgiving. Parts arrive out of spec. Conditions drift. Humans compensate with judgment and workarounds—but that reliance is now the bottleneck.

In many industrial systems, more than 50% of downtime and defects are attributable to manual handoffs, misinterpretation, or delayed human intervention. Systems that depend on constant human compensation do not scale or hold under stress.

Physical intelligence must understand physics and intent at the same time. It has to reason about forces, geometry, and constraints while inferring what a human operator is trying to accomplish when instructions are incomplete or conditions change. It must perceive noisy signals, decide under uncertainty, act through constrained machines, and verify outcomes in real time. Anything that cannot close that loop—without waiting on a human—stalls. Latency becomes downtime. Ambiguity becomes scrap.

Much of today’s “physical AI” stops at abstraction: models that describe the world but do not operate within it. The harder work is embedding intelligence directly into workflows where decisions carry cost, safety, and uptime consequences. The systems that matter start with concrete problems—dispatching labor, routing freight, coordinating people and machines—and fuse perception, decision-making, and action into loops that improve through use.

This is how intelligence becomes industrial: when it absorbs variability instead of escalating it, understands both physics and human intent, and compounds only by operating in the real world.

What unlocks venture-scale returns:

Physical intelligence platforms that own execution in the real world via frontier models and/or high-margin hardware, to augment or replace expensive labor, and become the operating layer for critical workflows.

What to avoid:

Analytics, copilots, or dashboards that observe industrial processes without controlling outcomes.

Emerging leaders:

Archetype AI*, GenLogs*, ReadyOn*, Skild, Physical Intelligence, General Intuition, Gecko Robotics, Covariant, Dexterity

6. The Industrial AI Operating Systems

30–40% of industrial project time is lost to coordination rather than work — not because decisions are wrong, but because execution is fragmented across systems that don’t share state.

A factory gets delayed because permits live in email, engineering drawings sit in shared drives, procurement is tracked in spreadsheets, and approvals move manually through inboxes. Field crews arrive without updated plans. Equipment sits idle waiting on signatures. Everyone is working — but the system isn’t. As stakeholders multiply, throughput collapses under the weight of handoffs. Timelines slip invisibly and capacity goes unused.

Industrial AI operating systems replace this fragility with software that owns execution.

AI agents operate against a single source of truth and actively advance work. They turn unstructured requirements into task graphs, sequence dependencies, route approvals, allocate crews, enforce compliance, and surface exceptions before they become failures. Humans intervene only where judgment or accountability is required.

In practice, workflows move continuously instead of episodically: permits assemble from live engineering data, procurement triggers when dependencies clear, crews dispatch against real-time constraints, and compliance artifacts generate as work progresses.

This is not analytics or assistance. It is execution infrastructure.

When software owns the workflow end-to-end, progress stops stalling in meetings and inboxes. Work advances by default. Projects that once took months of coordination compress into days.

Industrial capacity doesn’t just grow — it becomes usable.

What unlocks venture-scale returns:

Owning the orchestration layer that replaces document-driven, multi-constituent workflows with a single system of record for execution.

What to avoid:

Copilots, chat interfaces, or task-level agents that assist humans but do not control workflow state, sequencing, or outcomes.

Emerging leaders:

Euclid Power*, Heliux, Palantir

7. AI Native Industrial Roll-Ups

The next phase of industrial scale will not come from building a handful of pristine, clean-sheet factories. It will come from owning many imperfect ones—and systematically upgrading them.

The U.S. manufacturing base is dominated by small, profitable, under-automated plants operating in supply-constrained niches. They generate cash, but lack the labor, tooling, and systems to expand throughput or modernize operations. Capacity exists, but it cannot scale.

AI-native industrial roll-ups invert the traditional consolidation playbook. Instead of relying on leverage and cost cutting, they treat factories as deployment surfaces for a shared intelligence stack. Robotics, AI, and a standardized factory operating system are rolled out across similar sites, turning automation into reusable infrastructure rather than bespoke capex. Each acquisition expands the data set, improves the models, and lowers the cost of upgrading the next facility.

The result is compounding operational advantage. EBITDA expands through throughput, yield, uptime, and labor leverage—not financial engineering. Over time, these networks behave less like collections of plants and more like distributed industrial platforms: learning systems that get better with scale and restore capacity by upgrading what already exists.

What unlocks venture-scale returns:

A repeatable roll-up model where proprietary robotics coupled with AI reliably expand EBITDA across acquired assets, compounding operational advantage with each new facility.

What to avoid:

Financial roll-ups that rely on leverage and cost-cutting without changing how work is actually performed.

Our Commitment to America

We believe the modern American Frontier is created by companies that integrate software, hardware, manufacturing, and deployment into real systems, creating asymmetric capability that keeps America on the technological frontier.

And we at Venrock have been quietly investing in the American Frontier for years, before it had a name, a narrative, or any consensus behind it. It reflects the same conviction that guided earlier generations of our firm to back the infrastructure, energy, and computing systems that defined prior industrial eras. From aerospace to autonomy to quantum to defense systems, we leaned in when it was harder, slower, and out of step with prevailing venture incentives.

We thrive on supporting companies whose impact unfolds over long arcs—where conviction matters more than cadence, and where real systems take time to earn their place in the world.

A Call to Builders

This is not nostalgia or ideology.

It is a belief that free societies endure through pushing the frontier—the ability to build, deploy, and adapt faster than the challenges they face.

Building here is slower and harder than software—but that difficulty is the moat. Decisions are irreversible. Systems must work under regulation, scrutiny, and real-world constraints.

Much of the next era will still be built on screens. Software, models, and abstractions will continue to compound.

But some builders will shape the world more directly.

We back those working where software meets matter—where ideas must be embodied in systems that move, generate power, produce materials, defend nations, and endure real-world constraints. Where software becomes capability, and capability becomes resilience.

This is long-duration work.

This is the work of the next thirty years.

This is our calling.

Footnotes

*Venrock portfolio company

Critical materials defined as materials essential to energy and advanced manufacturing with 1. high supply-chain risk and 2. serves an essential function in more or more technologies. Examples include lithium (batteries), rare earth elements (magnets - neodymium), metallurgical coal for steelmaking (inclusive of anthracite), and copper (electrification).

The U.S. Department of the Interior published the final 2025 List of 60 Critical Minerals vital to the U.S. economy and national security: https://www.federalregister.gov/documents/2025/11/07/2025-19813/final-2025-list-of-critical-minerals

While everyone debates which model will achieve AGI first, the real constraint is hiding in plain sight: electricity. ChatGPT alone consumes 564 MWh daily—enough to power 18,000 homes. GPT-4 training required 50+ MW for months. The entire AI industry's power demand is projected to grow 8,000% as models scale and deployment accelerates.

The hyperscalers know this. Amazon just signed the largest corporate renewable energy deal in history—10.5 GW across 100+ projects. Google committed $20 billion to renewable infrastructure. Microsoft is restarting Three Mile Island's nuclear reactor and paying $16 billion over 20 years for the electricity. Meta is building dedicated solar farms larger than Manhattan just to power their AI training clusters.

These aren't sustainability initiatives—they're survival strategies. When your business model requires consuming small cities' worth of electricity, power availability becomes an existential constraint.

But here's what the hyperscalers discovered: throwing billions at renewable energy doesn't guarantee getting electricity. Despite $2.2 trillion in annual global renewable investment, 60% of projects face delays and 30% get cancelled outright. The bottleneck isn't technology or capital—it's operational chaos.

What the Industry Actually Needs

The renewable energy industry has built its business model around predictable processes designed for an industry scale that we left behind decades ago. To understand that scale of dysfunction, consider that the U.S. built 200 total power plants in 2005. In 2022, we built 10,000 solar projects—a 50x increase in complexity with zero infrastructure evolution to match.

Despite regulatory headwinds, renewable energy remains the cheapest and fastest source of new power (on track to deliver 93 percent of power buildout this year).

Each renewable project requires orchestrating dozens of stakeholders through interconnection queues, environmental reviews, tax credit compliance, and multi-million-dollar financing structures. When any single workflow breaks down, projects collapse. Of the $300 billion in annual U.S. renewable investment, $25-30 billion goes to "soft costs"—administrative overhead that creates the bottleneck.

Project delays cost an average of $2 million each. Cancellations cost $4 million. At current failure rates, that's $84 billion in annual waste from workflow dysfunction alone.

What the industry desperately needs isn't more capital or better technology—it's operational infrastructure that can handle complexity at scale. Specifically:

• Leverage AI to automate complex workflows. Manual coordination across dozens of stakeholders becomes impossible when project volume grows 50x. The solution isn't more project managers—it's eliminating the need for human coordination through intelligent automation.

• Continuous learning to improve outcomes. Every renewable project generates massive amounts of unstructured documentation: regulatory filings, engineering reports, financial models, compliance certifications. Converting this chaos into structured insights requires domain-specific AI trained on proprietary and industry-specific transaction data, not generic tools.

• Single Source of Truth Platforms. Multi-stakeholder transactions fail when information fragments across spreadsheets, email chains, and Box folders. Success requires centralizing all project data, workflows, and communication in collaborative environments that become mission-critical infrastructure.

Enter Euclid Power

Euclid Power built exactly what the renewable industry needs: an AI-powered operating system that transforms chaotic multi-stakeholder workflows into automated pipelines. But their full-stack, outcome-focused approach is what makes them inevitable winners.

AI-enabled Operating System

While most companies sell software and leave execution to customers, Euclid takes end-to-end responsibility for project outcomes. They don't just provide tools for compliance tracking—they deliver compliant projects. They don't offer document management—they manage the entire transaction from financing through completion. Customers pay for guaranteed results, not software licenses.

This outcome-based model creates unprecedented alignment. Euclid's success depends entirely on project success, and so their AI gets trained on what actually works, not what theoretically should work. The numbers prove this approach, growing exponentially over the last three years.

Founder-Market Inevitability

Jacob, Brian, and Ryan didn't just study this market—they built Goldman Sachs' renewable investing business from $0 to $3B in deployed projects. They lived through the operational chaos that kills deals and know exactly which workflows matter most. This isn't product-market fit—it's founder-market inevitability.

The Complete Solution Advantage

Euclid enters at the highest-stakes phase: project financing. But rather than stopping at transaction management, they become the operational backbone for everything that follows. Every successful project feeds their platform with proprietary data about what works in practice—pricing patterns, regulatory requirements, risk factors—creating intelligence that unlocks new opportunities and better outcomes.

Since its founding, Euclid has onboarded more than 12 gigawatts of solar and storage capacity, has supported over $10 billion in renewable energy investment, and is already embedded in more than 1,000 distributed generation (DG) and utility-scale projects.

Market Transformation Creates Massive Opportunity

Recent regulatory changes make AI-native platforms even more critical. Trump's "Big Beautiful Bill" creates new compliance requirements—supply chain audits, six-year data retention mandates, foreign component restrictions—with severe penalties for errors. Manual processes can't handle these requirements at scale.

What makes Euclid compelling is how they've transformed renewable project management where customers actively prefer paying them to handle critical workflows rather than just accessing tools. Their AI-powered platform leverages proprietary project data to deliver outcomes competitors can't match, establishing them as essential infrastructure for the industry's transition to abundant energy.

The AI Services Thesis

Taking a step back, this represents our AI services thesis:

• Services are the wedge into the high-touch, slow-change vertical

• Vertical-specific software enables those services to be delivered more efficiently

• AI supercharges the software, unlocking high margin and fast scale

• The solution is delivered via superior outcomes and product workflows (not chatbots)

• Proprietary customer and vertical data creates an AI moat of scalable expertise

In this market specifically, the customer buying behavior prefers full-service over pure software; and customers actively want to pay more for Euclid to handle the work rather than just access the tools.

This model effectively flips the typical tech-enabled services model—instead of using services to land software deals, they're using software to make services more efficient and sticky.

The Meta-Opportunity

The renewable energy industry faces a choice between AI-native platforms or administrative collapse. While AI companies race to build breakthrough models, they're all bottlenecked by the same constraint: getting critical electricity infrastructure projects from signed contracts to operational reality.

Companies like Euclid aren't just enabling clean energy—they're enabling AI's future. The window for AI-native platforms in critical infrastructure is wide open now, but it won't stay that way. The winners will be those who move first, build data moats fastest, and lock in operational workflows before the market catches up.

We are thrilled to have led Euclid Power’s Series A and to support them in delivering the infrastructure that powers the future of AI.

This isn't an accident—it's a carefully constructed economic moat that has insulated legal services from normal market forces for decades.

But all monopolies eventually face their reckoning. When the marginal cost of diligence, drafting, or research approaches zero, pricing based on time becomes indefensible. Every other industry where technology erased time as the primary cost driver—trading, advertising, parts of consulting—saw incumbents lose pricing power overnight. Law is the last great holdout.

AI is about to end it.

The Billable Hour's Last Stand

The traditional $900B legal services model relies on three foundational principles that AI fundamentally transforms:

• Price tied to inputs (hours worked, not outcomes delivered)

• Scale driven by associate leverage (junior lawyers billing at senior rates)

• Efficiency as enemy (faster work means lower revenue)

This structure has survived for decades because it benefits everyone except the client. Partners maximize profits, Associates get training, and firms avoid the risk of outcome-based pricing. The client pays for this elaborate theater of productivity.

Within each practice area, 30–60% of billable hours (estimates based on Associates vs Partner billable hours) go to repeatable, rules-based tasks—the exact work AI can compress by 50–90%. A $1.5B firm generating $500M in billables from tasks that AI can reduce to minutes faces an existential math problem. Even with conservative adoption, that's a $250M revenue line under direct pressure.

For legacy firms, AI creates an impossible choice:

• Adopt AI → hours fall → revenue falls

• Resist AI → clients defect to faster, cheaper competitors

Without pricing transformation, AI adoption creates a utilization crisis: too many associates for too little work.

Why BigLaw Can’t Pivot

• Revenue Recognition Shock — Fixed pricing collapses revenue per matter unless volume spikes — requiring a totally different sales motion.

• Partner Compensation Politics — Switching to flat fees sparks fights over how to split a smaller pie.

• Utilization Metrics Become Useless — Kill hours and you kill the KPI every dashboard, comp plan, and review depends on.

• Pricing Discipline Risk — Efficiency gains flow to the client, not the firm, when procurement demands discounts.

• Cultural Resistance — Billables are part of professional identity. Removing them is like removing the scoreboard.

Real BigLaw Innovation Leaders Are Seeing This Now: "We will eat into our write down buffer because fees are so high with how efficient we are. We need biz dev to be supercharged because we need to redeploy those hours or end up having to look at fixed fees." - Head of Innovation, Top 10 Firm.

Partnership Economics Make Long-Term Bets Impossible.

Most law firm partnerships distribute profits annually, leaving little capital for multi-year R&D investments. Even the most forward-thinking innovation programs stall when year-one ROI becomes mandatory.

Partnership Economics Make Innovation Impossible: "Partners want their PPP at the end of the day...that's a difficult conversation that requires a lot of matrix political conversations."

Enter the AI-Native Law Firm

An AI-native law firm isn't "BigLaw with Copilot." It's a completely different species that operates on software economics while delivering legal outcomes.

In an AI-native firm, legal outputs are delivered like software—fast, predictable, and with clear SLAs. For clients, elite-quality outcomes at a fraction of the cost. For founders, SaaS-like margins in a market that's been pricing like luxury consulting for centuries.

The Most Vulnerable Practice Areas

The industry has convinced clients that complexity requires armies of associates. In reality, most high-value legal work follows predictable patterns that AI can automate.

These represent $45B+ in annual fees across practice areas where the "complexity" is largely manufactured scarcity.

The AI Pricing Inversion

The traditional legal pricing model has historically aligned attorney time with client value—a framework that worked well in a manual-intensive environment. As automation transforms service delivery, this relationship is evolving toward outcome-based value recognition.

The market perception shift happening now:

• Old World: "We pay for the time of the best lawyers"

• New World: "We pay for guaranteed results at a predictable price"

When clients internalize this shift, hours become a liability—a signal of inefficiency. BigLaw's brand promise ("we have the most thorough people") becomes a red flag, while AI-native firms' promise ("faster, more accurately, for a fixed price") becomes the default.

The AI Pricing Advantage:

• Counter-Positioning: AI-native firms start with fixed fees and never give them up

• Cost-to-Serve Arbitrage: $50K matter costs $2K to deliver; the rest is margin

• Client Trust Flywheel: Predictable pricing → more volume → better models → higher margins

Here are a few examples where AI can collapse costs while improving results:

The Pattern: In each case, clients pay premium prices for time-intensive processes when they actually want binary outcomes or risk assessments. AI delivers the same (or better) outcomes at 95%+ cost reduction while providing faster, more accurate results.

The Hidden 10x Market Expansion

Traditional pricing models developed to serve complex, high-stakes matters, which naturally created market boundaries that AI can now help bridge

The Current Reality:

• BigLaw serves ~5,000 large enterprises with $50K+ annual legal spend

• 32 million SMBs and startups rely on solo practitioners or go without quality legal counsel

• Most high-stakes legal work (M&A, compliance, IP protection) remains exclusive to Fortune 500 companies

This isn't market selection—it's market suppression. The same legal expertise that costs $200K for a Fortune 500 company could be delivered for $20K to a mid-market company if you remove the artificial constraints of hourly billing.

The AI-Native Opportunity: When AI-driven workflows compress a $200K due diligence project to $20K, suddenly:

• Mid-market M&A explodes: 50,000+ companies can now afford professional-grade deal analysis

• Startup compliance scales: Every SaaS company can access always-on regulatory monitoring

• Global expansion democratizes: SMBs can navigate international tax and employment law

• IP protection universalizes: Individual inventors and small companies can file patents professionally

Market Math: Breaking down the 32M underserved businesses by segment:

• 10M mid-market businesses with complex legal needs: $5K average annual spend = $50B

• 22M smaller SMBs with basic but regular legal requirements: $1K average annual spend = $22B

• Total newly accessible market: $72B+ in revenue

This isn't market share transfer—it's market creation.

Three Paths Forward: Where Billion-Dollar Companies Will Be Built

AI will penetrate legal through three distinct models, but only the latter two offer venture-scale returns. For founders, the choice determines whether you're building a feature or dismantling an industry.

1. AI Tools for Incumbents (The Crowded Lane)

Examples: Harvey, Casetext CoCounsel, Spellbook

These are horizontal platforms that help BigLaw optimize their broken business model rather than replace it. They face structural limitations:

• Commoditized moats: Same foundation models = same results = race to zero margins

• Adoption ceiling: BigLaw's billable culture actively disincentivizes efficiency (5-15% actual usage despite licensing)

• Model dependency: OpenAI/Anthropic control your pricing and strategic roadmap

What BigLaw Innovation Leaders Actually Think: "I don't think [Harvey] understands the industry well enough to really think that you're going to be the platform for legal. Adding another [tool] with the primitive UI that they have and frankly not a good understanding of how to solve the problems...no one would use it." - Innovation Director, Top 20 Firm

For Founders: Unless you have a clear path to proprietary data or workflow ownership, you're building a feature that Microsoft will absorb. The incumbents don't want to be disrupted—they want to be optimized.

2. AI-Native Law Firms (The Category Killers)

Build law firms from day one where AI handles 80-90% of the work and humans focus on strategy, client relationships, and complex judgment calls. This isn't about making lawyers more efficient—it's about making lawyers unnecessary for most legal tasks.

Core capabilities:

• End-to-end automation of high-volume workflows

• Embedded AI agents running matters in parallel

• Domain-specific reasoning models on proprietary data

• Continuous learning loops improving speed/accuracy

• Cross-practice orchestration in a single workflow

Examples:

• M&A diligence platforms replacing entire junior associate teams.

• PE/VC Fund Financing & Support

• Always-on compliance firms for highly regulated industries.

• Vertical-Specific Contract Lifecycle Firm

• IP & Patent Prosecution Engine

• Real Estate Transaction and Title Clearance Firm

• Global Immigration-as-a-Platform

Why they work:

• Deliver in days what BigLaw takes weeks to complete

• 50-70% lower fees while maintaining higher margin

• Handle 10x more matters per lawyer

• Control the full stack: workflows, data, delivery, and client relationship.

• Transparent pricing, real-time updates, guaranteed outcomes

• Every matter improves the data moat and speeds future delivery.

• Counter-positioned against BigLaw’s billable model — pricing advantage they can’t match without destroying their economics.

3. Hybrid Service-Data Platforms (The Sleeper Opportunity)

The most interesting opportunity isn't just disrupting legal services—it's using legal services as a data collection mechanism to create entirely new markets.

These aren’t “law firms with a database.” They’re service engines designed to collect proprietary, high-value datasets as a byproduct of client work, then monetize that data in adjacent markets. They win when the secondary value of the dataset is as large — or larger — than the primary service revenue.

Examples:

• Patent & IP Intelligence — Service arm prosecutes patents; dataset powers a subscription analytics tool for R&D teams and investors. (Think PatSnap, but bootstrapped from a legal workflow.)

• Commercial Contract “Market Standard” Database — Reviews thousands of NDAs/MSAs; anonymized clause benchmarks sold to procurement and legal ops. (Like TermScout, but fueled by an active service pipeline.)

• Litigation Outcome Prediction — Litigation prep services feed a predictive engine sold to insurers and litigation funders. (Like Lex Machina with a live data feed.)

• Regulatory Compliance Map — Ongoing compliance monitoring builds a jurisdictional playbook sold to corporates. (Ascent RegTech-style, but client-verified.)

• Real Estate Deal Analytics — Transaction services generate permitting, title, and contract datasets for developers and lenders.

The data moat advantage:

• Proprietary by Design: Data only exists by doing the actual legal work—can't be scraped or purchased

• Real vs. Theoretical: Shows what actually happens in practice, not what's supposed to happen

• Continuously Updated: Gets better and more valuable with every new client engagement

• Impossible to Reverse Engineer: Competitors can't access the underlying legal work that generates the insights

Why this model is interesting:

• Dual revenue streams: Service funds data acquisition; data scales without linear costs

• Compounding moats: More clients = better data = better service = more clients

• Market expansion: Sell insights to much larger adjacent markets

Why the Window Is Now

The legal industry's isolation from market forces is finally breaking down. General Counsels face relentless cost pressure while BigLaw continues raising rates as if demand is infinite.

Once credible AI-native alternatives exist, GCs will move fast. Brand advantage decays when speed and price differ by 5–10x. Switching risk is low when contracts, filings, and negotiations are machine-tracked and instantly auditable. First movers get political capital for “modernizing legal” in the eyes of boards and CEOs.

The tipping point will be demand-led, not supply-led—clients will force the change.

Clients Are Forcing the Change: "We've had conversations with major clients that say to us, look. We expect AI to make things less expensive. You're either gonna figure that out or we're just gonna pay you 20% less next year, and you ****ing figure it out." - Innovation Director, Global Firm

The transition won't be without friction—data privacy concerns, regulatory adaptation, and client trust in AI for high-stakes work will create temporary barriers. But when the performance gap reaches 5-10x at 1/5 the cost, these concerns become implementation challenges, not adoption blockers.

The AI-Native Playbook

The window for AI-native law firms is wide open now, but it won't stay that way. Here's how to build one before the market catches up:

1. Regulatory Wedge — Launch in Arizona, Utah, UK, Australia where non-lawyer ownership is allowed

2. Beachhead Vertical — Dominate 1-2 high-automation areas before expanding

3. Data Moat — Capture proprietary datasets from every engagement

4. Pricing Disruption — Fixed pricing at 50% of BigLaw rates with 60%+ margins

5. Stack Control — Own orchestration layer, not just GPT APIs

6. Hybrid Lawyers — Those who combine legal reasoning, AI orchestration, and product thinking.

The Reality Check from Inside BigLaw: "I think we have two to three years, realistically, for [AI-native firms] to hit. The firms who are the legacy firms will move into the background. That's what I'm trying to do with [our firm] is not just make it future ready, but future proof." - Director of Practice Innovation, Top 50 firm.

The Endgame: A New Legal Universe

By 2035:

• Most transactional work fully automated.

• Fewer Associates, more strategic Partners.

• Output-based pricing becomes default.

• Proprietary legal playbooks guarded like trade secrets.

• Handful of AI-native firms control majority of repeatable workflows.

The next Cravath won't be a partnership of hundreds—it will be a platform company with global brand, billions in ARR, and SaaS-like margins.

This is not theory—the economic math makes it unavoidable. Once clients experience legal services that are 5–10x faster at 1/5 the price, they won't go back.

The winners will be those who:

1. Move first

2. Build data moats fastest

3. Lock in client workflows before the rest of the market wakes up.

The legal industry has operated as a protected monopoly for centuries. That protection is about to expire. If you're building in this space—or thinking about it—now is the time. The industry will not have another window this wide open in our lifetime.

This massive sector remains dominated by human-driven, low-tech businesses where titans have built their brands on historical reputation and prestige rather than measurable performance. They operate on centuries-old business models, relying on armies of highly paid professionals working through labor-intensive processes that have barely evolved since the typewriter gave way to the word processor.

But that is about to change.

A new wave of AI-native professional service firms is emerging, promising to resegment this $20T industry and relegate the legacy guard to history.

A New Fork in Professional Services History

While many focus on AI making existing software more efficient, the true revolution is happening as AI pushes directly into domains previously exclusive to human experts: strategic negotiations, creative problem-solving, and high-stakes decisions. These critical domains have remained stubbornly resistant to software automation until now.

We're witnessing a fork that will split the professional services landscape into two distinct futures:

Legacy Professional Services:

• Dominated by established incumbents with century-old processes

• Knowledge siloed by human experts with limited documentation

• Services bottlenecked by human cognitive and time constraints

• Premium pricing models based on artificial scarcity

Elite AI-Driven Professional Services:

• AI-native firms delivering demonstrably superior outcomes

• Expertise amplification across entire organizations

• Services that scale beyond traditional human constraints

• Value-based pricing tied to measurable outcomes

This transformation isn't about automating existing processes—it's about fundamentally expanding what's possible in professional services. We're witnessing the birth of the next generation of industry-defining brands, built from the ground up to harness AI's transformative potential.

The Most Vulnerable Service Verticals

While every professional services vertical will be impacted, some are particularly ripe for disruption based on their knowledge-intensive nature, rule-based workflows, and pattern-dependent decision-making.

Combined, these five leading incumbents alone represent over $240 billion in annual revenue and employ more than 900,000 professionals. And they're merely the tip of the iceberg in their respective industries, representing massive markets ripe for AI-powered disruption.

The subsectors towards the top generally involve more structured data processing and clear rule applications, with the lower still containing significant judgment elements but within well-defined frameworks.

The Five Principles of Elite AI-Powered Service Firms

These new AI-powered service firms operate on fundamentally different principles that combine business innovation with technical capabilities:

These principles fundamentally transform what professional services can be—delivering exponentially better outcomes while operating with software-like margins.

Why Clients Will Embrace AI-Native Professional Services

Why would clients take the risk of moving away from the elite, proven brands to these new emerging AI-native services? The compelling advantages across different market segments will quickly overcome initial hesitation:

For Consumers and Small Businesses

• Democratized Access to Expertise: AI-native firms offer enterprise-grade expertise to all clients regardless of size. Small businesses receive the same quality of advice previously exclusive to Fortune 500 companies.

• Transparent, Fixed Pricing: Predictable, fixed-fee pricing models replace opaque hourly billing, eliminating cost anxiety and making elite services accessible to broader markets.

• Objectively Superior Results: When a new tax advisory firm identifies 40% more deductions or a legal firm wins 30% more cases, reputation becomes secondary to measurable performance.

• Instant, 24/7 Responsiveness: Immediate, high-quality guidance at any hour creates stark differentiation from traditional firms with their scheduling delays and limited availability.

For Enterprise Clients

• Complete Verification Capabilities: AI-native firms provide full documentation of reasoning, allowing executives to verify quality rather than relying on brand reputation as a proxy.

• Comprehensive Analysis At Scale: When AI-native firms analyze 100% of relevant cases, regulations, or scenarios while legacy firms sample only 5-10%, the performance gap becomes undeniable.

• Integration with Modern Operations: As client businesses become increasingly digital, they'll prefer service providers whose operations integrate seamlessly with their systems rather than firms operating on century-old paradigms.

• Institutional Knowledge Continuity: AI-native firms offer knowledge that doesn't walk out the door when star practitioners leave, eliminating key person risk that plagues traditional firms.

Trust Accelerators

The transition will accelerate through:

• Human-AI Partnership that preserves client relationships while delivering superior cognitive capabilities

• Process Transparency transforms the traditional black box of professional services into a verifiable glass box

• Empirical Performance Data that makes the superiority of AI-native firms increasingly difficult to ignore

• Heritage Brand Partnerships where legacy firms adopt AI capabilities to create hybrid models

By 2030, AI-augmented service delivery will be the expected standard, with traditional human-only approaches viewed as expensive anachronisms in most professional domains.

Transforming the Margin Profile: From Services to Software Economics

The economics of professional services are about to undergo a fundamental transformation as AI redefines the cost structure of delivering expertise.

Legal Services

Insurance Services

Wealth Management

This transformation enables AI-powered service firms to deliver premium services at competitive prices while maintaining software-like margins—fundamentally disrupting traditional service economics and creating powerful incentives for rapid adoption across professional service verticals.

The Path Forward: A Once-in-a-Generation Opportunity

This transformation represents perhaps the largest opportunity in the AI landscape today. For entrepreneurs and investors, the $20 trillion market of pure human expertise is now accessible in ways previously unimaginable.

What makes this moment historic isn't just technology improvement—it's a fundamental expansion of human capability. Like software's transformation of manufacturing, AI will redefine what professional services can be, democratizing access to world-class expertise previously confined to elite firms serving the largest enterprises.

For investors, this represents a multi-trillion dollar opportunity to back the next generation of professional service firms—companies that will redefine excellence in their fields while operating with software-like economics.

The market for pure human expertise is about to be transformed—not replaced, but enhanced, expanded, and elevated to heights we're only beginning to imagine.

And it's just getting started.

The original Transformer paper solved a fundamental problem: how to process sequences without recurrence. DeepSeek showed us how to make the training of that architecture 90% cheaper.

Now, Titans tackles an equally fundamental challenge: how to give AI systems true memory capabilities. And it's not just another incremental improvement - it's a complete rethinking of how AI systems remember and process information. Here's why this matters and what it could unlock.

Why This Changes Everything

Today's most advanced language models, despite their impressive capabilities, suffer from a fundamental limitation: they can't truly remember. They can write a novel or debug your code, but they can't remember what they learned yesterday. Every conversation starts fresh. Every insight gets lost.

Current attempts to solve this are all workarounds:

• RAG (Retrieval-Augmented Generation) - an anti-pattern that is just fancy search

• Vector databases - glorified lookup tables

• Extended context windows - expensive band-aids

• Sliding window attention - losing most of the context

It's like trying to build a computer without RAM - you can do it, but why would you?

It’s one of the most fundamental limitations in the transformer architecture that is holding back AI today.

What Makes Titans Different

This isn't just another paper about making models bigger or training cheaper. Titans introduces three fundamental innovations that change the game:

1. True Test-Time Memory: The researchers introduce a neural long-term memory module that's fundamentally different from traditional approaches. Instead of just expanding context windows or adding retrieval mechanisms, Titans learns to memorize at test time.

• Learns what to remember while running

• Gets smarter with every interaction

• Maintains context across months of use

• Actually learns from experience

2. Surprise-Based Memory Instead of trying to remember everything (like RAG) or nothing (like standard Transformers), Titans uses a sophisticated "surprise" metric to determine what's worth remembering:

• Immediate surprise: "This is unexpected"

• Historical surprise: "This pattern is important"

• Contextual surprise: "This changes what we thought we knew"

3. Adaptive Forgetting The real breakthrough isn't just in remembering - it's in forgetting intelligently:

• Automatically identifies outdated information

• Preserves critical insights while pruning routine data

• Maintains memory efficiency at scale

The Technical Leap Forward

Just like DeepSeek showed us training could be 90% cheaper, Titans shows us AI systems can be 100x more capable through memory. Here's how:

1. Scale Beyond Imagination

• Handles sequences >2M tokens long

• Maintains linear computational scaling vs quadratic of transformers

• Processes more context than GPT-4 and Claude combined

2. Efficient Processing

• Parallelizable training algorithm

• Fast inference despite deep memory

• Intelligent memory consolidation

3. Architecture That Makes Sense Three variants for different needs:

• Memory as Context (MAC): For deep understanding

• Memory as Gate (MAG): For efficient processing

• Memory as Layer (MAL): For easy integration

What This Unlocks

When we think about what becomes possible when AI systems can maintain deep, persistent memory.. these aren't theoretical use cases - they're transformative applications enabled by Titans' specific technical capabilities.

1. The Developer That Never Forgets

"I see you're modifying the authentication system. We tried this OAuth approach in March - it caused rate limiting issues. Here's what happened, why it failed, and three better approaches based on what we learned."

2. The Analyst That Actually Learns

"This market pattern matches what we saw before the 2023 banking stress, but with three critical differences. Here's why this time is different, based on analyzing every similar situation in the past decade."

3. The Research Assistant That Builds Understanding

"While these symptoms suggest bronchitis, I'm seeing a pattern in the patient's history that matches 47 similar cases where the initial diagnosis was wrong. Here's the complete analysis and what we learned from those cases."

Why This Is the Next Transformer Moment

Just as Transformers solved sequence processing, Titans solves memory. This means:

1. True Learning Systems

• Systems that actually improve with use

• Memory that persists and evolves

• Understanding that deepens over time

2. Exponential Value Creation

• Each interaction makes the system smarter

• Knowledge compounds over time

• Insights transfer across domains

3. Fundamental Architecture Shift

• From stateless to stateful AI

• From pattern matching to true learning

• From short-term to long-term understanding

What's Left to Prove

The path from breakthrough research to production AI is well-worn at this point. With Titans, three key things need to be proven before this becomes more than an exciting paper:

1. Memory Economics: While the paper shows linear computational scaling (vs quadratic for transformers), the real costs are still unclear:

• How much more expensive is training?

• What's the GPU memory overhead in production?

• Can smaller companies afford to deploy this?

2. Memory Reliability: Test-time learning is exciting, but raises hard questions:

• How stable is the memory over weeks or months of operation?

• What happens when the memory gets corrupted?

• How do you safely update or reset memory states?

3. Production Readiness: The paper shows promising results on research benchmarks, but enterprise deployment is another story:

• How does memory work in distributed systems?

• What's the story for backup and recovery?

• How do you monitor and debug memory issues?

None of these challenges seem fundamental - they're engineering problems, not research problems. But they're the difference between an exciting paper and a technology that transforms how we build AI systems.

The good news? The AI infrastructure ecosystem is far more mature than when Transformers were introduced. We have the tools, talent, and incentives to solve these challenges. It's not a question of if, but when.

Looking Forward

This isn't just another improvement - it's a fundamental shift in how AI systems work. Just as Transformers enabled the current AI boom, Titans enables the next wave: AI systems that truly learn, remember, and grow.

The next generation of billion-dollar AI companies won't be built on bigger models - they'll be built on better memory. The future belongs to systems that don't just process information, but actually learn from experience. Where real ROI can be delivered by:

• Solving memory problems that directly tie to revenue or cost reduction (real ROI in year one)

• Building deep moats through accumulated context (the value should compound over time)

• Focusing on sectors where context and historical knowledge have enormous value (enterprise knowledge, software development, r&d, operations)

The companies that understand this shift now will define the next decade of AI. The race to build truly intelligent systems just got a lot more interesting.

If you're building in this space, we'd love to hear from you. The future of AI isn't just about bigger models - it's about systems that truly remember and learn.

In one weekend, history changed. We were steady as a freight train on a pre-DeepSeek timeline, dominated by a few companies with access to vast computational resources, ready to invest $500B in a historic (and somewhat comical) AI infrastructure. To a post-DeepSeek fork, proof that efficiency in training and inference can yield superior results, and that innovation is not a result of unfettered access to resources but a creative response to constraints. Everyone had to step back and ask how did this happen? Why did it come out of China and not the US? While some see this as a challenge to American AI leadership, I see it as validation of why we partnered with Manos, Oussama and team many months back, and in introducing Oumi to the world, the world's first unconditionally open AI lab.

Why Oumi Is America's Strategic Imperative

Oumi isn't just another AI company – it's a fundamental reimagining of how we develop AI and America's opportunity to be the true stewards of frontier AI development. Working with academics from over 13 universities, they've built what has shaped into becoming the "Linux of AI" – a platform that ensures every future "DeepSeek moment" comes from American soil, embodies democratic values, and is unconditionally open.

Here's why this matters:

1. Collective Innovation Beats Siloed Development: DeepSeek's breakthrough came from rethinking fundamental assumptions about AI training. Imagine what's possible when we have thousands of American researchers and developers able to experiment, iterate, and build upon each other's work in real-time. Oumi's platform makes this possible, enabling experimentation and collaboration at a scale no single company could match.

2. Democratic Values in Code: While recent breakthroughs raise questions about AI development ethics and potential biases, Oumi represents a transparent, collaborative approach aligned with democratic values. Their platform ensures that frontier AI development remains open, auditable, and aligned with Western principles of scientific openness.

3. America's Infrastructure Advantage: The United States possesses the world's most sophisticated AI infrastructure. What we've lacked is a platform that allows us to use these resources efficiently and collaboratively. Oumi's platform turns America's computational advantage into an unstoppable collaborative force.

History Repeats: The Linux Moment for AI

We've seen this story before. When Unix dominated computing, major tech players developed their proprietary, closed-source versions, claiming they were the only ones with the talent and resources to advance the technology. Then Linux changed everything – showing how open source could create something better, more versatile, and more trusted than any closed alternative.

AI is having its Linux moment, and Oumi is leading the charge. Just as Linux became the foundation for modern computing, powering everything from cloud infrastructure to mobile devices, truly open AI will become the foundation for humanity's next great technological leap.

True Open Source Means Everything is Open

The AI community has seen various interpretations of "open source" – from companies releasing model weights without training code to those providing limited access to their development process. But true open source means something more fundamental: a complete platform that enables open collaboration and innovation.

Oumi stands apart by providing a truly comprehensive open platform that lets you:

• Build at Any Scale: Train and fine-tune models from 10M to 405B parameters using cutting-edge techniques like SFT, LoRA, QLoRA, and DPO

• Maximum Flexibility: Work with text and multimodal models (Llama, Qwen, Phi), synthesize training data with LLM judges, and deploy efficiently with vLLM and SGLang

• True Openness: Seamlessly integrate with both open models and commercial APIs (OpenAI, Anthropic, Vertex AI), with no vendor lock-in

• Universal Access: Run anywhere—from laptops to clusters to any cloud—with comprehensive evaluation across standard benchmarks

This isn't about checking boxes or partial transparency – it's about creating a truly open ecosystem where innovation can flourish. When breakthroughs happen, whether from established players or newcomers, the entire community should have the tools needed to build upon them. Most importantly, it's about giving researchers and developers the foundation they need to skip the plumbing and focus on pushing the boundaries of what's possible.

Building the Foundation for Frontier AI Innovation

What makes Oumi unique is its unprecedented collaboration with leading academic institutions. The platform has been developed in partnership with researchers from Massachusetts Institute of Technology (MIT), Stanford University, University of California, Berkeley, Carnegie Mellon University, Georgia Institute of Technology, University of Illinois Urbana-Champaign, University of Washington, California Institute of Technology, University of Michigan, Princeton University, Harvard University, Columbia University and University of Texas at Austin.

This coalition of America's top AI research institutions isn't just for show – it's a fundamental rethinking of how we advance frontier AI development. Each institution brings unique expertise and perspective to the platform, from Berkeley's pioneering work in reinforcement learning to MIT's advances in neural architecture search.

Accelerating Collaborative Innovation

What makes Oumi transformative is how it enables researchers to build on each other's work. When breakthroughs like DeepSeek's emerge, dozens of research teams naturally want to understand and improve upon them. With Oumi's platform:

• Teams can immediately begin experimenting with new approaches

• Researchers can collaborate across institutions in real-time

• Successful innovations can be quickly verified and built upon

• The entire community benefits from shared progress

This isn't just about sharing code – it's about creating a complete environment for frontier AI development where every breakthrough catalyzes the next. The next major AI innovation won't just come from a single lab or company—it will emerge from the collective effort of America's best researchers and developers working together on an open platform.

The Power of American-Led Open Source

Linux didn't just become the world's dominant operating system because it was free – it won because it was better. The same dynamics that made Linux successful are even more powerful in AI:

• Trust and Transparency: Organizations need AI they can trust with their sensitive data and intellectual property. Oumi's open source approach provides the transparency needed for proper security auditing and compliance while protecting American interests.

• Flexibility for Innovation: Different industries and applications need different AI capabilities. Oumi's platform allows organizations to customize and optimize AI models for their specific needs, rather than being locked into one-size-fits-all solutions.

• Efficiency at Scale: By enabling collaborative development and efficient resource sharing, Oumi's approach can dramatically reduce the cost of AI development while accelerating progress.

Securing America's AI Leadership

DeepSeek's breakthrough isn't a threat to American AI leadership – it's an opportunity to demonstrate why our approach to innovation is superior. The future of AI won't be determined by who has the most GPUs, but by who can innovate most effectively and responsibly. America's advantages in this new paradigm are overwhelming:

• The world's most sophisticated AI infrastructure

• The largest concentration of AI research talent

• A culture of open source collaboration

• Strong intellectual property protections

• A robust venture capital ecosystem

Oumi's platform leverages all these advantages, turning America's computational resources and talent pool into an unstoppable engine for AI innovation.

This Is Our Moment

We've partnered with Oumi because this is America's moment to define the future of AI. Their platform isn't just a technical achievement – it's a strategic imperative for ensuring that AI development remains transparent, ethical, and beneficial to society as a whole.

The next chapter in AI development won't be written by a single company or country. It will be written by thousands of researchers, developers, and organizations working together on open platforms. But make no mistake – this collaboration will be led by American innovation, powered by American infrastructure, and guided by American values.

This is how America wins the AI race – not by building walls, but by building bridges. Not through secrecy, but through openness that builds on our strengths. Not through control, but through empowerment.

The future of AI is open. The future of AI is American. And it starts with Oumi.

Join us in building this future. Start with pip install oumi, or check out Oumi's platform, join their Discord community, or contribute to their GitHub repositories.

DeekSeep r1 represents a significant leap in AI – bringing an order of magnitude improvement in cost efficiency and accessibility at frontier AI performance. It’s a fork in history in AI. But, it likely could only be built now, not a year or two ago, given its extensive reliance on prior investments and distillations from existing leading frontier models.

Let’s remind ourselves what makes DeepSeek unique:

DeepSeek r1 Overview:

• Cost Efficiency: r1 is 93% cheaper to use per API call compared to o1, making it highly cost-effective.

• Local Deployment: Can be run locally on a high-end workstation with 37 GB of RAM (1 GB per 1B active parameters in FP8).

• Performance: No rate limits observed, and batching/compute optimization further reduces costs and increases token speed.

• Geopolitical Context: Released shortly after "Stargate," hinting at geopolitical dynamics in AI development.

Real:

• Market Success: #1 download in its App Store category, surpassing ChatGPT, Gemini, and Claude.

• Quality: Comparable to o1 in quality but lags behind o3.

• Algorithmic Breakthroughs: Innovations like FP8 training, MLA, and multi-token prediction significantly improved efficiency in training and inference.

• Training Cost: Verified $6M training cost (compute only), though this figure is misleading without context.

• Hardware Innovation: Novel architecture using PCI-Express for scaling.

Nuance:

• Hidden Costs: The $6M figure excludes prior research, ablation experiments, and access to large GPU clusters (e.g., 10k A100s). Real cost is much higher for labs without existing infrastructure.

• Distillation Reliance: Likely relied on distillation from GPT-4o and o1, raising questions about export restrictions on GPUs versus access to advanced models.

• Geopolitical Implications: Highlights the inefficacy of restricting GPU exports while allowing access to distilled models, undermining the purpose of such restrictions.

Short & Long Term Winners & Losers

Short Term

Losers:

• Nvidia: Market scare due to potential decreased demand for high-end GPUs for training.

• OpenAI / Anthropic / Google: Pressure to reduce costs or accelerate innovation to compete with more efficient models.

• Stargate: $500B looks a bit out of touch now

• Cloud Providers (AWS, Azure, GCP): Customers could Immediately demand for the hyperscalers to offer DeepSeek or similarly cost-efficient + performant models.

• Overall Training Capex: This trend is not favorable for near-term capital expenditure in AI training or the "power" theme.

Winners:

• AI at large: The research in AI just accelerated by 2 - 3 years.

• Open Source First: Gains from community-driven innovation and model adaptation.

• New Model Developers: Easier entry with less need for extensive hardware.

• Coding Agents & Assistants: Efficiency of complex tasks enables faster, cheaper handling of complex coding tasks, boosting productivity through parallel computing.

• Academic Researchers: More accessible high-performance AI models for research.

• Investors in open source AI companies: Potential high returns from investments in cost-effective AI solutions.

• Benefit for AI Users: Companies that use AI (software, internet, etc.) stand to gain significantly.

Long Term

Losers:

• China: Lack of comparative AI data center infrastructure will throttle their ability to maximize and keep up.

• OpenAI: Possible decline in proprietary model value if open-source alternatives thrive.

• Traditional Software Companies: Risk of losing market share to AI-enabled competitors at a more accelerated pace.

• Data Labeling Companies: As AI models become more efficient and require less labeled data for training, companies specializing in data annotation (e.g., Scale AI) could see reduced demand.

Winners:

• ASI: We will get to ASI (super intelligence) much sooner than we anticipated.

• US Dominance in AI: Leveraging US strengths in computing power, innovation, and infrastructure to more efficient and performant models will lead to AI escape velocity.

• Enterprises: More AI, everywhere.

• Explosion of New Use Cases and Specialized Models: Innovation in sector-specific applications.

• Nvidia: Long-term demand for inference GPUs, new markets in hardware optimization.

• Healthcare: AI-driven improvements in complex areas such as Healthcare for diagnostics, treatment, and efficiency.

Additional Consideration:

• Capacity Utilization: Even though DeepSeek and similar technologies make model training more efficient, the trend of models growing in size and complexity means we are likely to continue maxing out available computing capacity. Just two years ago, models were significantly less capable than today's, yet we've seen no slowdown in the demand for computational resources. This suggests that while efficiency helps, the hunger for more advanced, larger models will keep pushing the limits of what's available, potentially leading to new cycles of infrastructure investment or innovative approaches to capacity management.

• Nvidia (Short-Term Loser): While Nvidia may face short-term uncertainty, its dominance in GPU manufacturing and its ability to pivot to inference-focused hardware could make it a long-term winner.

• Cloud Providers: While they may face short-term pressure, hyperscalers like AWS, Azure, and GCP could adapt by offering DeepSeek-like models as part of their services, turning a potential threat into an opportunity.

And for good reason. DeepSeek just shook the AI world. Making us all fundamentally rethink how we develop AI going forward.

And of course the spiraling narratives start; calling for the end of the AI Bubble, the end of Nvidia’s ascent, the rise of China AI, and the slowdown in growth. Nasdaq futures already down Sunday night “DeepSeek Shakes Up Stocks as Traders Fear for US Tech Leadership”

But I am calling BS on it all.

In reality, this is the best thing that could have happened and further reinforces why the US will dominate the AI race.

A New Fork in History

On one hand, DeepSeek represents an incredible breakthrough in performance and cost-efficiency. DeepSeek-R1 and DeepSeek-R1-Zero have set a new standard for AI development at a fraction of the cost.

On the other hand, it’s not short of controversy—questions around potential model development ethics, actual development costs, and biases towards Chinese government-controlled narratives.

But it doesn’t matter. What DeepSeek just demonstrated is a fork in history.

From a pre-DeepSeek timeline:

• Dominated by a few companies with access to vast computational resources, leading to high cost, high barrier models.

• A narrative was one of increasing computational power and data size dictating AI progress.

To a post-DeepSeek timeline:

• Instead of a race for ever-larger models, a belief efficiency in training and inference can yield superior results.

• Where AI development focuses more on algorithmic innovation and less on brute-force computation.

• Innovation not as a result of unfettered access to resources but a creative response to constraints, possibly leading to more sustainable tech development practices.

First, how DeepSeek is different from other models

DeepSeek’s approach to AI stands out in several key ways from OpenAI, Google, or Anthropic.

1. Pure Reinforcement Learning (RL)

• Traditional Models: Most AI models, including OpenAI’s GPT series, rely heavily on supervised fine-tuning (SFT), where humans provide labeled data to teach the model.

• DeepSeek’s Innovation: DeepSeek-R1-Zero is the first model trained entirely via reinforcement learning (RL) without SFT. This allows the model to self-evolve its reasoning capabilities through trial and error, leading to behaviors like self-reflection and multi-step error correction.

• In simple terms, most AI models start by memorizing examples (like a student copying answers). DeepSeek R1 skips this and learns entirely through trial and error, like a robot figuring out how to walk by experimenting. It tries solutions, gets feedback on what worked, and improves over time.

2. Group Relative Policy Optimization (GRPO)

• Traditional Models: Training often requires a separate "critic" model to judge responses, which is computationally expensive.

• DeepSeek’s Innovation: GRPO eliminates the need for a critic model by comparing batches of answers at once and rewarding the best ones. This reduces costs and speeds up training while maintaining high performance.

3. Self-Improvement Loop

• Traditional Models: These models depend on human-generated data for learning, which limits scalability.

• DeepSeek’s Innovation: DeepSeek-R1 generates its own practice problems and learns from its mistakes, creating a self-improvement loop. This allows the model to scale learning without being bottlenecked by human input.

4. Internal Monologue on Full Display

• Traditional Models: Models like ChatGPT often operate as "black boxes," making it difficult to understand their decision-making processes.

• DeepSeek’s Innovation: DeepSeek-R1 incorporates explainable AI tools, displaying its internal reasoning steps during problem-solving. You can follow along it’s logic chains to understand the eventual output.

Second, the importance of high-quality training data

One of the key factors behind DeepSeek’s success is the high-quality training data it started with. While many AI models rely on vast amounts of uncurated or noisy data, DeepSeek took a different approach:

1. Curated "Cold Start" Data

• Before diving into full reinforcement learning, DeepSeek-R1 was initialized with a small but highly curated dataset of 80,000 reasoning chains. These examples were carefully selected to ensure clarity, accuracy, and logical structure.

• Why this matters: This "cold start" phase provided the model with a strong foundation in reasoning and problem-solving, enabling it to generate coherent and logical outputs from the outset. Requiring orders of magnitude less training data to start.

2. Focus on Reasoning and Logic

• Unlike models trained on general-purpose datasets, DeepSeek’s initial data emphasized reasoning, logic, and structured thinking. This focus allowed the model to excel in tasks requiring multi-step problem-solving, such as math, coding, and scientific research.

• Why this matters: By starting with high-quality, reasoning-focused data, DeepSeek avoided the pitfalls of models that struggle with logical consistency or fail to "show their work."

Why the U.S. Wins in AI Thanks to These Innovations

It’s simple. We have what the rest of the world doesn’t – the most sophisticated, purpose-built AI compute clusters in the world. DeepSeek represents the fork in history we needed to use that infrastructure 100x more effectively.

1. Cheaper, Faster Training - 90% Cheaper, or 10x Faster.

• DeepSeek’s Innovation: Group Relative Policy Optimization (GRPO) eliminates the need for a separate "teacher" model, reducing training costs by up to 90%.

• Why the U.S. Wins: By lowering training costs, GRPO allows U.S. companies and research institutions to train more models with the same resources. This efficiency could lead to a surge in AI development, growing model performance by 10x from today with the same compute.

2. Self-Improvement Loop on Steroids - 10x More Effective Data Utilization

• DeepSeek’s Innovation: The self-improvement loop allows models to generate their own practice problems and learn through trial and error, scaling efficiently with massive computing power.

• Why the U.S. Wins: With our vast network of data centers and Nvidia's progress, we can generate billions of synthetic problems and solutions in days. This accelerates learning without the bottleneck of human input, enabling faster and more efficient model training. Models could generate 10x more training data through reasoning and understanding with the same data sets.

3. Better Use of Compute = Better Models = 10x More Inference Time

• DeepSeek’s Innovation: Redirects wasted computing power toward productive tasks like self-play and distillation, creating smarter, more efficient models.

• Why the U.S. Wins: We redirect time and resources from massive training compute clusters to train larger and more complex models to massive inference clusters focused on solving real problems, from drug discovery to climate modeling, rather than wasting energy on inefficient training processes.

US strengths about to go on display

The creative innovations by DeepSeek + the US strengths in computing power, innovation, and infrastructure can lead to 90% cheaper training costs, 10x faster training, with models that could utilize data 10x more effectively and spend 10x more time in inference than training.

It’s about to get exciting.