The AI Chokepoint Framework

The constraint is not wafer starts. It is advanced packaging. TSMC's CoWoS lines run full, and the overflow that reaches everyone else is second-tier work: industry capacity is projected near 200k wafers per month by end-2026 with TSMC at ~120k, leaving roughly 80k to be split as overflow while the large GPU packages stay in-house.

What makes this category rank first is that the restraint may be deliberate. A physical constraint yields to capital eventually. A policy choice does not. If TSMC is pacing expansion for reasons of Taiwanese industrial stability rather than capability, no amount of money relieves it — and that is the load-bearing claim worth testing against actual capex guidance before repeating it.

The longest replacement cycle outside EUV. GE Vernova closed Q2 2026 with 116 GW of gas power equipment backlog and slot reservations against roughly 20 GW of annual output, and is already taking reservations for 2031 delivery. Across the three OEMs, lead times stretch toward eight years.

Two things temper it. Only 53 GW of that 116 is firm equipment backlog — the rest is paid options that have not converted. And GE Vernova's own executives have said turbines are not what is gating data centre buildouts. If the real gate is permits and interconnection, the backlog converts more slowly than the headline implies.

Which is why the second seat holds electrons rather than equipment. Existing interconnection cannot be manufactured at any price.

The acute pinch today and the most likely to unwind first. SK hynix holds roughly 57–58% of the HBM market and posted a 76% operating margin in Q2 2026 as HBM4 mass production began.

The instructive fact is what the market did with those numbers: the stock fell nearly 9% on the print, near a 52-week low, as investors weighed pricing durability. Record margins were sold. All three makers are expanding into the shortage simultaneously, and 2026 capex is targeted at the high KRW 40T range. A low multiple on peak cyclical earnings is a warning, not a bargain.

Hence a royalty model in the third seat. When the cycle turns, IP royalties compress far less than fab margins.

The chokepoint is CUDA, not the silicon — and it is the one constraint here that is actively eroding. Triton enables write-once GPU programming, torch.compile abstracts hardware, and ROCm 7 sits within 10–30% of CUDA on most workloads. System-level integration still holds training at scale, but Nvidia's own behaviour tells the story: equity into CoreWeave, open-sourcing Dynamo and Nemotron, licensing Groq for inference. That is a company widening a moat because the original one is loosening.

Meanwhile the escape route compounds faster than the incumbent. Broadcom's Q1 FY2026 AI revenue rose 106% year on year to $8.4bn, against a $73bn backlog and a $100bn 2027 target, with custom ASIC shipments growing 44.6% versus 16.1% for merchant GPUs.

The third seat is the one indifferent to the outcome. Every accelerator in this fight — merchant or custom — is designed on EDA tools with no leading-edge substitute.

The chokepoint here is launch, and it is tightening rather than easing. SpaceX is phasing out Falcon 9 for the unproven Starship while New Glenn and Vulcan fly irregularly and Neutron is not yet on the pad.

SpaceX itself became investable on 12 June 2026 — the largest IPO in history, $135 per share, roughly $75bn raised at a valuation near $1.8 trillion. Its February acquisition of xAI pivoted the company explicitly toward orbital data centres.

The thesis deserves scepticism it rarely gets. Varda calculates orbital compute at roughly 3x the cost per watt of terrestrial. Near-term deployment means two 8kW satellites in 2027 against a stated 5GW ambition — about six orders of magnitude. And the honest observation is that if orbital compute works, it flows through categories 05 and 06 anyway: the satellites still need accelerators, memory and laser links. Only power and cooling leave the planet.

Categories 01–07 are entirely supply-side. This one exists because a supply-only framework cannot test its own premise.

Every chokepoint above is priced off these three companies' capital budgets. The load-bearing assumption — that new data centre capacity gets absorbed by inference workloads and earns an adequate return — is made here, not in the supply chain. If it fails, it fails here first and everything upstream reprices.

Worth noting what the disclosed record shows: the legends are not buying the chokepoints, they are buying the buyers. Berkshire added roughly $17bn of Alphabet in Q2 2026; Pershing Square holds Microsoft and Amazon as top-four positions while avoiding Nvidia, Tesla and Apple entirely. Neither owns a single name from categories 01–07.

The weakest category on the test, and the one most likely to matter in five years. Inference moving on-device and into physical systems is structurally real; the revenue is not yet. This fails on inelastic demand for the plain reason that the demand has not arrived.

It is included as a positioned watch-list rather than a conviction basket — the same position Orbit occupied before the SpaceX listing changed its arithmetic.