Former White House AI advisor David Sacks frames the infrastructure push as the economy itself. Dean W. Ball argues the diffusion controls — the US export rules on advanced AI chips meant to protect that lead — are eating into the only window where the buildout math holds.
The United States is mid-pivot on AI. One arm of government is underwriting what former White House AI advisor David Sacks calls "the economy" itself: a trillion-dollar infrastructure buildout of data centers, accelerators, and power. Another arm is writing rules that shrink the customer list those data centers are being built to serve. The contradiction is not a policy failure. It is structural, and it is tightening.
Dean W. Ball, a policy commentator with prior US-administration AI ties, made the contradiction concrete in a Hyperdimensional essay quoted by Simon Willison on June 26. Ball's frame is two clocks. The first is the recoup window. Training a frontier model, meaning one of the most capable systems currently in development, costs a great deal of money, and the model's economics only work if a meaningful share of that cost is recovered in a narrow post-release window. That window ends when a competitor ships a better model and the lab's offering drops from frontier to sub-frontier, still useful but no longer state-of-the-art. The second clock is the global TAM, or total addressable market. Nobody finances a $100 billion data center, Ball argues, to serve the roughly one hundred companies US diffusion rules would currently permit. The thesis only pencils out if US AI services can sell to a global customer base.
That is where the policy contradiction sits. The capex case and the access case pull the same revenue base in opposite directions.
Sacks set the capex stakes when he told Fortune in May that AI was already responsible for 75 percent of US GDP growth in the first quarter, and likely to keep delivering. Morgan Stanley has put numbers around that claim: its latest hyperscaler capex forecast for Amazon, Alphabet, Meta, Microsoft, and Oracle is roughly $805 billion in 2026 and around $1 trillion in 2027, framed as a 2.5 to 3 percent tailwind to US growth. Those numbers are forecasts, not realized spend, but they reflect what the major cloud and AI platforms have publicly committed to building.
The access case runs through Washington's AI diffusion controls, the export rule machinery administered by the Commerce Department and refined under the Trump administration's executive-order framework. These rules restrict which countries and which end users can receive advanced AI chips, and increasingly the model weights and services that sit on top. The stated intent is national security: keep frontier compute out of the hands of adversaries. The side effect, Ball argues, is to shorten the recoup window by reducing the customer base, and to invalidate the global TAM assumption by regulatory fiat.
Three pressure points follow.
First, the recoup math. Every week a frontier model sits under a diffusion constraint is a week its developer cannot sell to the customers the data center was built for. If the post-release window is already narrow before sub-frontier competition arrives, restricting access during that window converts a recoverable investment into a stranded one. Ball's claim is not that diffusion rules destroy the AI economy. It is that they compress it, which at the margins amounts to the same thing.
Second, the buildout math. Ball puts the per-facility cost at $100 billion or more, a figure that only pencils out if the operator can serve customers across most major markets. If the served market shrinks to a regulatory perimeter, even one defined generously as all non-adversary nations, the unit economics on the marginal megawatt deteriorate. The hyperscalers do not need every customer, but they need enough of them to justify the next ten campuses.
Third, the policy posture. The White House cannot simultaneously be the world's loudest champion of AI capex and the architect of a customer list that excludes most of the world. The two positions are not adjacent. They are inconsistent. Ball's contribution is not to take a side but to name the inconsistency and ask what gets cut first: the access rules, the capex targets, or the assumption that both can coexist.
The honest answer is that nobody in Washington has yet had to choose. The diffusion rules remain narrow enough, and the global customer base for US AI services remains wide enough, that the two clocks keep ticking past each other. That tolerance is shrinking. As frontier model training costs continue to rise and sub-frontier catch-up cycles continue to compress, the recoup window gets shorter. As the diffusion rule set expands to cover model weights and downstream services, not just chips, the customer list gets smaller. At some point the two clocks stop being asynchronous.
The structural question Ball is asking is therefore not whether US AI policy is right or wrong. It is whether the United States can keep building trillion-dollar infrastructure for an industry whose customers its own rules are designed to exclude. The buildout and the diffusion controls are pointed at the same revenue base. So far they have been aimed in opposite directions with enough slack that nothing has broken. The slack is the part the policy has not yet priced in.