Analysis
The Trump administration is weighing a second round of semiconductor tariffs that would, for the first time, reach past bare chips into the finished products built around them -- laptops, data-center servers, and gaming consoles -- according to multiple reports citing people familiar with the internal discussions. The plans remain preliminary and fluid, with any tariffs likely phased in rather than applied all at once, but the direction of travel is clear: the cost of importing the physical infrastructure AI runs on is going up, not down.
In January, the administration imposed a 25% tariff on certain AI chips, with carve-outs for data centers, R&D, startups and consumer devices. Commerce officials have told industry contacts in private conversations that those exemptions may not survive the next round. In their place, the administration is reportedly floating a framework that ties duty-free import allowances to a company's own US fabrication investment -- the more capacity a company commits to building domestically, the more it can import tariff-free in the meantime. That turns tariff exposure into a direct, negotiable function of onshoring speed rather than a flat tax.
Who actually eats this
The exposure is broad and uneven:
- Nvidia and AMD -- both sell the GPUs at the center of every AI buildout; a tariff on finished servers effectively taxes the boxes their chips ship inside, on top of whatever hits the chips directly
- Dell, HP and the server OEMs -- assemble and import the physical machines hyperscalers and neoclouds buy by the thousand
- Microsoft, Sony and console makers -- Xbox and PlayStation hardware would be swept in as "finished products containing semiconductors," alongside every laptop sold in the US
- AWS, Azure, Google Cloud and every VC-backed neocloud -- CoreWeave, Lambda and their peers finance GPU-backed debt against a cost basis that a new tariff would move mid-loan
None of this is happening in a vacuum. The same week the tariff story surfaced, OpenAI published its first independently verified benchmarks for Jalapeño, the custom inference chip it originally unveiled in June built with Broadcom -- Pulse covered the unveiling here. SemiAnalysis engineers visited OpenAI's labs and ran the workloads themselves using InferenceX, finding 1.5 to 1.9 times more AI work per watt and 1.7 to 3.6 times lower latency than Nvidia's Blackwell GB200/GB300 systems, at roughly half the power draw. The comparison has real caveats -- Jalapeño uses newer HBM4 memory, so the fairer fight is against Nvidia's upcoming Rubin chips, and Jalapeño only handles inference, not training, with volume production not until 2027. But a company building its own chip specifically to control its own supply chain and cost curve is exactly the kind of hedge a tariff-uncertain 2026 rewards.
A third, unrelated data point landed the same week: Stripe and buyout firm Advent abandoned their pursuit of PayPal, walking away from a $53 billion offer after PayPal's stock ran roughly 40% ahead of the bid on a strong quarter and takeover speculation. Tariffs had nothing to do with that deal falling apart. But stacked together -- a tariff threat raising the cost of physical AI infrastructure, a frontier lab racing to build proprietary chips to escape exactly that kind of exposure, and a $53 billion fintech buyout that simply couldn't get done on its original terms -- the pattern for founders and GPs is the same: assume the deal, and the infrastructure under it, costs more and takes longer to close than it did six months ago.
For infrastructure-heavy portfolio companies, the practical takeaway is diligence-specific: any GPU lease, data-center build, or hardware-heavy term sheet signed today should model a scenario where import costs rise mid-contract, not just at signing. A lease with no cost-escalation clause tied to trade policy is underwriting a risk the lessor isn't pricing in.
The counterweight is that this has happened before without landing. The administration floated similarly broad tariff expansions earlier in 2026 and walked several back after industry pushback; the January round's carve-outs for startups and data centers were themselves a concession to exactly the kind of lobbying now underway again. It's entirely possible this second round ships narrower than reported, or with new exemptions of its own. But even the credible threat of it changes underwriting behavior now -- and Jalapeño's benchmark reveal this same week is proof at least one company isn't waiting to find out.
