And Vijay Says…

America has a commanding lead in AI compute. Whether that becomes a lasting economic advantage is a separate question.

There is a growing argument that America loses the AI race to China if we don’t build data centers faster. I think there is a real concern behind it. Compute is a constraint, power takes years to build, and China is investing heavily in energy and industry.

But building more data centers is not a strategy by itself. So, are we measuring the wrong things? Partly, yes. We keep count of capacity: chips, megawatts, data centers. We count results far less.

I find it easier to think about three separate questions.

1. How much capacity does America need?

2. Who pays for it?

3. And how well can our economy turn it into productive activity?

The third one gets the least attention in the public debate.

America is ahead. By a lot.

The Lowy Institute’s November analysis concludes that America is winning. It estimates that US AI data centers have roughly 40 times China’s computing power and US AI chips 31 times, although the report doesn’t spell out how those multiples were built. American models draw more than 90% of combined web traffic to US and Chinese AI services. Lowy also says its model rankings are consistent with claims that the US is six to nine months ahead, a figure it cites but does not measure itself.

That strengthens the case for keeping our compute capacity growing.

Now put two numbers from the 2026 Stanford AI Index next to each other. US private AI investment in 2025 was $285.9 billion, against $12.4 billion in China. And in March 2026 the best American model led the best Chinese model by 2.7% on the benchmark Stanford highlights.

These aren’t like-for-like. Chinese state funding isn’t in the private investment figures, a benchmark covers a narrow slice of what models do, and capital spending pays back over years.

Still, it makes me ask how much capability and value we get from each additional dollar invested.

China’s progress doesn’t show compute is unimportant. Chip export controls have made it a real constraint for Chinese developers. But Chinese labs have shown that efficiency can close part of the gap. We should care about how well we use compute, not only how much we have.

Open models

China is strong in open-weight models. Per Lowy, Chinese models were downloaded about 1.8 billion times on Hugging Face in the third quarter of 2026, against 500 million for US models. That is 78% of the two countries’ combined total.

Downloads aren’t deployments, though. Many are small models and automated pulls, and American models still dominate consumer use.

Lowy also reports that Chinese models handled 55% of text requests on OpenRouter during one week in late October, against 38% for American models. That is one developer marketplace for one week, and Lowy itself notes that most people reach AI through apps and websites, where US models dominate. It is one signal of where developers are choosing to build.

A company that can run an open model on its own hardware and tune it may get good results at much lower cost than a large proprietary model would give. That is worth supporting, along with independent testing of provenance, security and behavior.

Neither open weights nor a model’s country of origin tells you whether it can be trusted.

The real constraint is speed to power

Most large training clusters don’t need to sit near customers. Some inference does. Developers already look at remote sites and their own generation.

The bottleneck is how fast they can get power, and new transmission in advanced economies can take four to eight years.

FERC moved in June, ordering six regional grid operators to justify or reform their rules for connecting large loads. The first responses are due November 16. Those will be proposals. The real test is whether connection times fall once new rules are in force.

In the meantime, dedicated generation may be the fastest option for some projects, even if it is more expensive or less environmentally attractive. Modernizing the grid is necessary, but it won’t solve the next two years of demand.

I’d like to see published connection timelines, utilities saying what each new large load requires, and a transparent split of incremental costs. Grid upgrades help manufacturing too, so we shouldn’t treat them as a subsidy for AI.

Water gets a lot of attention. Closed-loop cooling cuts water use but can raise electricity use, so the trade-off has to be checked site by site. In some places the power bill will matter more. In others, water availability will be the bigger concern.

Who pays if the forecasts are wrong?

Companies are committing hundreds of billions of dollars based on forecasts of demand, prices and hardware performance. Some of those forecasts will be wrong.

Cheaper inference could shrink the compute each task needs, or it could grow total demand. Both are plausible.

If a utility builds for a data center that gets cancelled, or that uses half the power it forecast, existing customers shouldn’t pick up the bill.

Ohio’s large-load tariff uses minimum payments. State commissions should look at long-term commitments, financial guarantees and exit charges, without piling on so much that good projects die.

Investors should keep the upside and carry a fair share of the downside.

The constraint nobody is measuring

Gartner expects cloud spending on running AI models ($23.3 billion) to pass spending on training them ($19 billion) this year.

The 2026 McKinsey survey says 80% of respondents report personal productivity gains, but only 37% report any positive effect on enterprise EBIT, and about 6% qualify as high performers, attributing at least 5% of EBIT to AI with significant overall impact.

The last two numbers are flat from 2025, even though adoption and spending went up.

Spending on running models is climbing, yet the share of surveyed companies reporting an enterprise profit effect isn’t. That doesn’t prove infrastructure is being overbuilt. Returns may be lagging the investment, or may be concentrated in a small group of companies. Either way, making intelligence more available isn’t automatically producing broader financial returns.

McKinsey also found that nearly three-quarters of high performers rebuilt workflows around AI, against about a quarter of everyone else. The survey can’t show cause, but it matches kind of what I’ve seen in enterprise work.

A bank automates document review and leaves credit approval alone. An insurer speeds up claims and keeps decision rights where they were. The AI works. The organization hasn’t changed enough to capture the value.

That is the argument in Migrating Scarcity. When intelligence gets abundant, scarcity moves to the next constraint, which is our ability to absorb it, authorize its use and redesign work around it.

Is China better at this? I don’t have the evidence to say. Stanford reports that China installed about 295,000 industrial robots in 2024, 54% of the world total, against about 34,000 in the US. That mostly reflects the size of its manufacturing base, but it is a reminder that deployment matters as much as model quality.

What I’d check in 2028

Enterprise returns. Move the share reporting any EBIT gain from current baseline of 37% that McKinsey reports to say 50%, and double the high performers from 6% to say 12%. These are targets I’d use to judge progress, not predictions. McKinsey’s sample is global, so I’d also want a US-only cut.

Grid timing. A one-third cut in the median time from a complete application to energization, measured against a published baseline. Alongside it, publish cancellations and any costs shifted onto existing customers.

Cost per task. Not the price per token. The cost of finishing a defined set of tasks at fixed quality, reliability and security, at least halved from today. This needs to be baselined quickly – I am surprised this hasn’t been done yet.

New things. New products, markets and companies that couldn’t have existed before AI. This is harder to count, but I’d start with whether business formation in AI-related sectors is growing faster than business formation overall.

What would worry me

If we miss the enterprise-returns targets while capacity keeps expanding, our ability to use AI isn’t keeping up with our ability to produce it. That wouldn’t necessarily mean we built too much. But it would make me question where the next dollar of investment should go.

I’d also want to know whether data center owners are earning adequate returns. Enterprise productivity alone doesn’t guarantee they made good investments.

America should build the compute it needs, modernize the grid and protect ratepayers. None of that removes the need to change the institutions that will use the intelligence.

As I argued in AI and the Next American Economy, we also need new products and industries, not only cheaper versions of today’s work.

Compute is the lead we already have. Using it productively is how we keep it.

Can America create economic opportunities as fast as AI transforms existing work?

The midterm elections have given Democrats control of both chambers of Congress while Republicans retain the White House. I expect plenty of political disagreements over the next two years. But I would hope both sides can agree on one objective: maximizing economic growth and productive employment.

AI gives us an extraordinary opportunity to do both. But there is a question I don’t think we are paying enough attention to.

What happens if AI eliminates existing work faster than the economy creates valuable new work?

Productivity doesn’t guarantee employment

The classical economic argument is compelling. Technology lowers costs. Competition lowers prices. Purchasing power increases, creating demand for other goods and services. New businesses emerge and employment follows.

I think AI will create these benefits too. But there is a complication.

Suppose AI makes software development ten times more productive. We could create ten times more software and entirely new applications. But AI might also develop, test and maintain most of that additional software.

We could end up with dramatically more output without a corresponding increase in employment.

History gives us plenty of examples of technology creating new work. It doesn’t prove that new work will emerge fast enough this time, especially when AI can automate much of that new work too.

I think the real economic challenge is to expand the frontier of valuable human activity faster than AI reduces the labor required for existing activities.

An early signal worth taking seriously

Stanford researchers Brynjolfsson, Chandar and Chen found that employment among workers aged 22 to 25 in highly AI-exposed occupations was approximately 19% below where it would have been if it had grown at the same rate as employment among less-exposed peers.

The difference was primarily associated with reduced hiring. The researchers also found no widespread economy-wide displacement and cautioned that the evidence doesn’t establish AI as the cause. Still, I think this deserves serious attention.

Work produces two things: economic output and human capability.

Junior engineers write code. Junior accountants reconcile accounts. Analysts build models. They gradually develop the judgment needed to handle more difficult work.

AI can remove those learning opportunities along with the work.

An individual company may rationally hire fewer juniors. But if enough companies do that, where will the next generation of experienced professionals come from?

This is something I explored in The Corner Case. A company can buy experienced talent instead of developing it. An entire industry cannot keep doing that forever.

We need more exploration, not just exploitation

Companies have strong incentives to use AI to make their existing operations more efficient. And they absolutely should capture those productivity gains.

But I think the bigger opportunity is in discovering economic activities that were previously impossible or uneconomic.

AI could make personalized education, preventive healthcare and entirely new types of scientific research commercially viable. The problem is that cheaper technology alone doesn’t create new businesses.

A startup might build an excellent application at a fraction of the historical cost but still struggle to reach customers. The years of litigation between Apple and Epic illustrate how valuable control of distribution can be, even when building software becomes cheaper.

New businesses still need capital, infrastructure, customers and permission to operate.

And there is another problem. Markets may underinvest in exploration when the benefits spill over to competitors or when the returns are too uncertain for conventional financing. We need established businesses to become more productive. We also need new businesses to discover what else becomes possible.

Three things I would ask Congress to do

1. Make creating and financing new businesses easier.

Reduce unnecessary regulatory fragmentation, improve market access and remove infrastructure bottlenecks, especially in electricity and transmission.

Fund foundational research that benefits multiple industries. Experiment with SBIR-style awards and prize challenges to encourage commercialization of new ideas.

I would rather see government help create more businesses than keep subsidizing the same large companies.

2. Invest in developing human capability.

I would propose an illustrative five-year, $2 billion apprenticeship experiment supporting 40,000 paid developmental positions annually, with public funding capped at $10,000 per position.

Test employer incentives against worker-directed funding. Require actual paid work, employer contributions and portable skills. Use randomized or phased enrollment to find out whether these programs create additional jobs or simply subsidize hiring that would have happened anyway.

There is good reason to be skeptical of subsidies. A 2026 NBER study found no detectable effect on hiring, earnings or retention from the Work Opportunity Tax Credit in Wisconsin. The authors identified eligibility-screening concerns and organizational inertia as possible explanations.

A better-designed program might work. We should test that assumption rather than take it on faith.

Congress should decide upfront what an acceptable cost per additional sustained job looks like. If an approach doesn’t deliver, stop funding it.

The Senate HELP and House Education and Workforce committees could develop the pilot, with funding through appropriations.

This addresses the entry-level pipeline. Displaced mid-career workers need attention too, but that’s a different problem requiring different interventions.

3. Reexamine incentives for investing in people versus technology.

Employers pay payroll taxes on wages, while qualifying equipment and certain software investments can benefit from 100% first-year depreciation.

That doesn’t automatically mean the tax system favors automation. Labor costs are deductible too, and employees may bear some of the economic burden of employer payroll taxes through lower wages.

But I would like Congress to examine whether targeted payroll-tax offsets for additional developmental employment could improve hiring without discouraging automation. Any such offset should be funded elsewhere so Social Security and Medicare are unaffected.

We should make technology cheaper to deploy. We should also make it economically attractive to develop people.

Why I still care about employment

There is a reasonable counterargument here.

If AI can produce more output with fewer workers, why not distribute the purchasing power and let people work less?

Maybe that will eventually be part of the answer.

But employment does more than generate income. It develops capability, creates opportunities for advancement and gives people a way to participate productively in the economy.

I don’t think we should preserve obsolete jobs. I certainly don’t think maximizing hours worked is a sensible goal either.

What I would like to maximize is the opportunity for people to contribute economically, develop valuable capabilities and earn higher real incomes.

What should we measure by 2028?

I would like Congress to establish an annual AI and Economic Opportunity report that tracks productivity, real wages, new employer businesses and early-career employment.

The apprenticeship experiment should have predefined measures of success, including cost per additional sustained job. We could evaluate implementation in 2028 and longer-term employment outcomes in 2029 and 2030.

For the broader economy, I would keep asking three questions.

Are we creating more productive businesses?

Are more Americans getting opportunities to develop valuable capabilities?

Are real earnings rising alongside productivity?

If AI dramatically increases output and profits while opportunities for productive employment keep shrinking, I would find it hard to call that a complete economic success.

We should be aiming for all three.

AI is making code cheaper to produce. That however doesn’t make software cheaper to own.

AI has dramatically increased the amount of code an engineer can produce. It hasn’t necessarily increased the amount of software an organization can safely own. Agents implement features, write tests, review their own work and fix the defects they find. So why should anyone read every line?

In Migrating Scarcity, I argued that when technology makes something abundant, scarcity moves to the next constraint. Code production is becoming abundant. The ability to establish that code is correct, safe and worth deploying is not expanding at the same rate. Code review needs to be rethought around that gap.

Passing tests is not the same as being correct

Let’s say we have an agent that implements a pricing change. It writes the code, generates the tests, runs static analysis and reports that everything passes.

Should a human read the diff? Maybe not. But someone still has to confirm the pricing rules are right. The implementation may faithfully execute a specification that left out an important business exception. The tests may cover every requirement in that specification. The agent may review its own work and find nothing wrong.

Everything passes, and the business still loses money.

Automated verification is only as good as the properties it checks. A type checker establishes certain facts about a program. A test suite shows that specific behaviors work under specific conditions. Neither shows that the right problem was solved. Automate verification wherever you can, but someone still has to decide what correct means. That decision matters more as code gets cheaper.

Three questions that decide where humans belong

Most organizations still run code review as a uniform ritual: someone writes, someone reads, the pull request merges. It was never a precise way to spend engineering judgment, and AI makes that obvious. A thousand lines of generated boilerplate may carry almost no risk. A five-line change to an authorization rule can expose customer data. Reviewing both the same way makes no sense.

I use three questions.

What happens if this is wrong? A defect in an internal prototype is not a defect in a payment system. Reversibility, blast radius and security exposure matter far more than line count.

Can correctness be verified independently? Some changes have clear contracts, strong tests and well-defined invariants. Others involve ambiguous requirements or business rules that are hard to state completely.

Who can recover the system when it breaks? A change can be correct today and still become a liability if nobody understands it well enough to diagnose it during an incident.

The principle that follows: the greater the consequence of failure, the stronger the independent evidence we should demand. Human attention belongs wherever that evidence is insufficient, whether that means reviewing code, specifications, architecture or operational readiness.

The 2 AM test

Imagine an agent builds a new subsystem. Every test passes, performance is excellent, security scans are clean. Six months later it fails at 2 AM.

Can the team explain the system’s critical assumptions, isolate the failure and recover service, even if the agents that built it are unavailable?

Nobody needs to understand every line. But if the team can’t operate the system without the agent that wrote it, what did the organization actually acquire? Cheap code with expensive ownership. In my native language Malayalam – we say you just bought an expensive dog that bites you 🙂

AI reviewing AI is not necessarily independent evidence

Suppose one agent writes the code, another writes the tests, a third reviews the implementation and a fourth fixes what the third finds. That sounds thorough.

But different agents are not necessarily independent reviewers. They may share training biases, assumptions or misunderstandings of the specification. Four checks can still miss the same error.

Agent review is valuable and should be used heavily. But count the confidence your checks justify, not the number of checks performed. Independent specifications, adversarial tests, production observation, fault injection and explicit business invariants can provide stronger evidence than asking a second model whether the first one did a good job.

The honest objection: human review was already unreliable

Plenty of human review is just rubber-stamping. Nobody reads a 3,000-line diff carefully. Approval is often a formality.

That is true, and it strengthens the case for reallocation. Reading everything produces the appearance of assurance. Reading the right things, with the authority to stop a release, produces the real thing. Humans should be reviewing specifications, business invariants, security boundaries, architecture and recoverability, and periodically sampling the automated process to catch failures in the process itself.

Accountability belongs here too. Every production change needs an accountable owner who accepts the evidence for deploying it and knows how the system will be operated. Removing manual review doesn’t remove that responsibility.

Who gets to learn?

Code review was never a perfect training mechanism, but it was one way engineers met unfamiliar code, questioned design choices and learned how a system fit together. If senior engineers only review the hardest problems and agents handle the rest, junior engineers lose the path to becoming the people who can handle those problems.

Watching agents succeed does not build judgment, and preserving low-value review rituals just to keep juniors busy is no answer either. Judgment comes from making decisions and living with the consequences. That means owning bounded changes, predicting failure modes before seeing an agent’s analysis, joining production diagnostics, and explaining why an implementation satisfies a business requirement.

This is no longer just a code review problem. Who gets the difficult work? How do we establish readiness for greater authority? Who pays to develop expertise when AI produces the immediate output more cheaply? Without deliberate answers, productivity can improve while engineering capability erodes.

The next scarcity

A company can have excellent automated tests and poor architecture. It can have secure code that nobody knows how to recover. It can have strong senior engineers today and no credible way to develop their successors. Reading every diff solves none of these, and neither does eliminating review.

The irony is that AI makes engineering judgment more important while potentially removing the work through which engineers learned to develop it.

The goal is not less human review. It is putting human judgment where it changes the outcome, and being able to say why the rest can be trusted.

Stop measuring how much code humans review. Start asking what justifies trusting the code they don’t.