This is arguably the hardest manufacturing on Earth, and it depends on a startlingly short chain of irreplaceable companies. Scarcity here is structural, not temporary — and the same forces that make the incumbents' moats historic have quietly closed the door on new entrants.
Silicon makes the hardware everything above it runs on. Operating fleets of that hardware — renting it out by the GPU-hour or the token — is the next layer up, Data Centers & Compute. This layer is the factory floor: chip designers, memory makers, foundries, packagers, equipment builders, and the networking that stitches accelerators into single machines.
The layer is best understood as a chain within the chain: designers (NVIDIA and its CUDA ecosystem, AMD, and the hyperscalers' custom parts) → high-bandwidth memory (SK hynix, Samsung, Micron) → foundries (TSMC, which fabricates essentially all leading-edge AI silicon and is the most geopolitically sensitive node in the whole stack) → advanced packaging (CoWoS, a real gate on output) → equipment (ASML, the world's only maker of EUV lithography machines) → networking (NVLink, InfiniBand, optics).
Value accrues at the choke points: a monopoly (ASML), a near-monopoly (TSMC), and a dominant ecosystem (NVIDIA plus CUDA). Every dollar spent anywhere in the AI stack eventually clears through one of these gates.
How few companies each step of AI hardware actually passes through — one concentration figure per node of the chain.
One node is moving against the concentration story and the chart will say so: SK hynix's HBM revenue share slipped from 62% to 57% in a single quarter (Q2 to Q3 2025, Counterpoint Research) as Samsung and Micron qualified HBM3E.
The hardest audited anchor for the figure: TechInsights counted NVIDIA at 98% of 2023 data-center GPU shipments and revenue. The other nodes hold the same shape — ASML is the sole EUV supplier, TSMC fabricates essentially all leading-edge AI silicon, and three memory makers split HBM between them.
Silicon takes roughly $0.22 of the dollar — the second-largest slice — and it is priced by structural scarcity. The compute layer above it can negotiate rates and shop across clouds; there is no shopping around ASML, TSMC, or CUDA. Choke-point suppliers set terms, and the rest of the stack pays them.
Illustrative split — assumptions in Methodology.
Hold both of these together, because both are true. The incumbents' moats are among the strongest in industrial history — a lithography monopoly, a foundry near-monopoly, a software ecosystem two decades deep. And the startup door is closed: the same performance demands, manufacturing precision, and supply-chain leverage that built those moats are now beyond what any new entrant can assemble. That tension is this page's question.
Chamath Palihapitiya helped get Groq off the ground in 2015. On December 24, 2025, NVIDIA closed its largest deal ever with the company — and the structure matters: roughly $20 billion in cash for most of Groq's chip assets, plus a non-exclusive license to Groq's LPU inference technology, plus hiring about 80% of Groq's staff, including CEO Jonathan Ross. Neither "acquired" nor "licensed" alone describes it. Chamath's own conclusion, stated on All-In: he will not invest in or incubate anything in this layer now. The performance demands are too high, the manufacturing precision too complex, and the supply-chain influence needed to secure adjacent components like memory is no longer possible for a startup.
"Lots of capital will be wasted chasing Groq and Cerebras' success."
Both companies, in his telling, made sense a decade ago — when these constraints were much more modest. The exit that looks like validation is actually the closing of the door.
AI compute has been growing roughly 3x per year, and Dwarkesh Patel's decomposition of that number is the most important arithmetic in the stack: about 1.4x from Moore's Law, about 1.2x from new fab construction, and about 1.8x from AI taking leading-edge wafer share away from phones and other devices. Fab growth is bottlenecked through at least 2030 by ASML's EUV tool supply. And the 1.8x term is the one that runs out: per SemiAnalysis, AI is just under 60% of TSMC's N3 node output in 2026, projected to reach 86% in 2027 — at which point there is nothing left to reallocate. The wall arrives around the end of 2027.
Gavin Baker calls HBM DRAM per flop the single most important lever for increasing token output per unit of compute — which makes the memory tier a choke point in its own right. But honesty requires the counter-annotation: HBM is the one node moving against the concentration story. SK hynix's revenue share fell from 62% to 57% in a single quarter as Samsung and Micron qualified HBM3E, and the battle now pivots to HBM4. Concentration here is loosening, not tightening.
NVIDIA is no longer just selling chips. Baker characterizes its financing structures as a "credit wrapper": third parties lend to GPU buyers while NVIDIA takes a revenue share above a price floor — effectively a royalty business layered on the hardware. Add equity stakes in nearly every major lab, and the company is becoming a tax on the whole stack rather than a vendor to it. Meanwhile its forward P/E sits near 22x, a multiyear low — the lowest since 2019 — which Baker reads as the market pricing in over-earning. He is skeptical of that pricing.
The layer's biggest risk is not technological, it is geographic. Essentially all leading-edge AI silicon is fabricated in Taiwan, export controls partition the market, and a US-China split runs directly through the supply chain. On China's domestic DUV progress, Baker's framing is a phase transition that still leaves a propeller-versus-jet gap against EUV — roughly a 25-year learning-by-doing deficit — and he argues the market overreacted to it. The slower erosion to watch is hyperscaler custom silicon: Google's TPUs and Amazon's Trainium chip away at merchant GPU margins from inside the biggest buyers. Those hyperscalers are covered where they live, on the Data Centers & Compute page — but their design ambitions are felt here.
98% of 2023 data-center GPU shipments (TechInsights) and the CUDA moat — now adding financing structures and equity stakes across the stack.
Fabricates essentially all leading-edge AI silicon — and is the most geopolitically sensitive single node in the entire stack.
The world's only maker of EUV machines. Its production rate bottlenecks global fab growth through 2030 and beyond.
Memory per flop is the key token-output lever — and the one node where concentration is loosening as all three qualify each new generation.
The credible second source for accelerators — competitive silicon still working against the gravity of the CUDA ecosystem.
The startups that proved the door is closed: both made sense a decade ago; one exited into NVIDIA in the largest deal the acquirer ever made.
The chain that makes frontier chips is a museum of monopolies: one lithography supplier, one leading-edge foundry, a handful of memory makers — and leading-edge capacity sold out years in advance. Groq and Cerebras got in when those constraints were a fraction of today's — that window is shut. The sharper point: even the gatekeepers are tolls with capped growth. The equity story in silicon belongs to whoever's revenue scales with demand, and this cycle produced exactly one of those.