Export Controls Just Created Their Own Workaround
Sakana AI's new Fugu model openly targets the gap left by Anthropic's Mythos export control suspension. We connect that story to BofA's hawkish rate pivot, the Fed's quiet bill buying, and the widenin
Bank of America flipped its 2026 rate call this week. The desk that had argued for a hold through year end is now penciling in three quarter point hikes, September, October, and December, lifting the funds rate to a 4.25 to 4.50 range. The reasoning leans on June’s FOMC projections, where nine of nineteen officials saw at least one hike by year end, and on new Fed Chair Kevin Warsh’s first press conference, which landed more hawkish than the market had priced. BofA economist Aditya Bhave summed up the call in five words: “the data call for hikes.” Markets, for context, are still pricing something closer to one move by early 2027, roughly 41 basis points of tightening for the year against BofA’s 75.
We would push back on the premise before the policy. Hiking into a government carrying $39 trillion in debt and roughly $1.56 trillion in annual interest expense does not behave like the textbook tightening of 1994 or 2004. Every 25 basis point move on that stock of debt is closer to $39 billion in additional annual interest income transferred from the Treasury to the private holders of that debt: money market funds, banks, foreign central banks, retail savers parked in bills. That is not contraction. That is the transmission mechanism running in reverse of what a hiking cycle is supposed to accomplish, and it is the same interest income channel we laid out at length in our working paper on fiscal dominance. If BofA’s hawkish pivot plays out, it becomes a live test of that thesis rather than a rebuttal of it.
There is a second, quieter piece sitting underneath this. Since mid December the Fed has been buying Treasury bills in the secondary market through what it calls Reserve Management Purchases, roughly $40 billion a month, alongside reinvestment of maturing mortgage backed securities principal back into bills. The Fed insists this is not quantitative easing, it is balance sheet maintenance for an ample reserves regime. The balance sheet does not care what it is called. It is growing, week over week, by design. Combine a Fed that is structurally a buyer of government paper most months with a hiking cycle that pays more interest on that paper to whoever holds it, and the inflationary mechanics do not point in the direction BofA’s headline implies.
On the technology side, watch what just happened with Anthropic. The US government issued an export control directive in mid June ordering Anthropic to suspend all access to its Mythos 5 and Fable 5 models for any foreign national, anywhere, including the company’s own foreign employees, over a disputed jailbreak finding. Anthropic complied and disabled both models for every customer rather than attempt to segment by nationality. Within days, Sakana AI, the Tokyo lab known for evolutionary model merging, launched Fugu: a multi agent orchestration system sitting behind a single API that coordinates a pool of other frontier models, including instances of itself, to deliver what it calls “frontier capability without the risk of export controls.” Sakana’s own launch language names Fable and Mythos directly. The pitch is sovereignty: route around any single government’s leash on any single lab by orchestrating across many. Whether or not Fugu’s benchmarks hold up under scrutiny, the positioning is the real story. Export controls aimed at individual frontier models may simply accelerate a shift toward orchestration layers that nobody can export control, because no single model inside the system is doing the controlling.
That story sits on top of a capex structure that is starting to look strained on its own terms. The major hyperscalers, Microsoft, Amazon, Alphabet, and Meta, are on track to spend somewhere between $650 and $725 billion combined on AI infrastructure this year, against free cash flow that is compressing across the board: Amazon’s trailing free cash flow down roughly 95 percent, Alphabet’s down 47 percent in a single quarter. A meaningful share of that demand is circular. Nvidia commits $100 billion to OpenAI, which spends it on Nvidia chips. Oracle signs $300 billion with OpenAI for data centers it fills with Nvidia silicon. Nvidia takes a stake in CoreWeave and commits billions to buy CoreWeave’s unsold capacity, capacity built on Nvidia GPUs. Microsoft owns roughly 27 percent of OpenAI and is its primary cloud landlord, recycling Azure revenue into the same chips it sells compute against. Two cash burning customers, OpenAI and Anthropic, now underwrite something like 1.05 trillion of a roughly $2.1 trillion revenue backlog sitting on the books of the four largest cloud providers. None of that is illegal or even unusual for an infrastructure buildout. It is, however, the same shape as the financing daisy chains that defined the back half of the 2000 cycle, and it is showing up exactly where you would expect it to: in the names that built the last three years of index returns.
That shows up directly in our own tracker. Microsoft sits at negative 24.04 percent year to date inside the MEGA9 basket, Meta at negative 14.58, Tesla at negative 9.93, Amazon barely positive at 0.85. These are not the speculative small caps taking the hit. These are four of the largest single commitments inside the basket. Microsoft’s own quarterly capex actually came down even as the OpenAI exclusivity arrangement that anchored its AI narrative came apart, and the market punished both the pullback and the unwind in the same week. Meta raised its 2026 capex guide to a $25 to $145 billion range and lost roughly 10 percent of its market value on the print. Nvidia, Apple, Alphabet, and Broadcom are still carrying the basket positive, up 11.88, 9.25, 11.15, and 13.30 percent respectively, but the dispersion opening up inside what traded as a single momentum block for three years is the real story of this stretch. The market is starting to price hyperscaler capex the way it should have from the start: as a bet with a payback period, not a multiple expansion story.
