One build-out, many layers — and wildly unequal economics
Every AI accelerator is a stack: a compute die, a CPU, memory, networking, and the instruction-set IP underneath. These six occupy different rungs, and the gap in economics is the first thing the cross-section reveals. NVIDIA takes ~90% of accelerator spend at a 71% gross margin on $215.9B of revenue. Arm, selling only IP, runs a ~98% gross margin but on just $4.92B. Micron, supplying the scarce HBM, swings from a $5.8B loss to record profit in two years. And Intel, once the center of the industry, is a spectator at <1% of AI accelerators and a 15-year revenue low. Same build-out; six completely different ways to be exposed to it.
The real question: does the CUDA moat survive custom silicon?
The debate the whole cluster orbits is whether NVIDIA’s ~90% share is durable. NVIDIA’s teardown leans yes — a 20-year CUDA software lock-in, full-stack systems and scarce packaging supply. But Broadcom is the live counter-argument made flesh: it co-designs custom accelerators (XPUs) for the exact hyperscalers that are NVIDIA’s biggest customers, and guides its AI business past $100B. Read side by side, the tension is explicit — NVIDIA’s four largest direct customers are ~61% of a quarter, and those same customers are funding Broadcom to build them an alternative. The cross-cut frames the bet cleanly: NVIDIA owns the merchant market, Broadcom owns the in-house one, and the question is which pool grows faster.
Monopoly economics show up as gross margin
The clearest signal across the six is that gross margin is a near-perfect map of pricing power. The two firms selling the least substitutable things — NVIDIA’s full-stack GPU and Arm’s architecture tax — post the fattest margins (71% and ~98%). The commodity layer, memory, runs a brutal boom-bust cycle even at the AI peak: Micron is structurally an oligopoly price-taker, which is why a record $37.4B year still trades on cyclicality fears. And the broad supplier, Intel, can’t hold margin at all (~39%, below every peer). The lesson the cluster makes concrete: in a capex super-cycle, where you sit in the stack determines whether you keep the dollar or just pass it through.
Intel is what falling off the leading edge looks like
Every other study here is a way to win the AI build-out. Intel is the cautionary tale of missing it. The company that defined the industry now cancelled its AI GPU, ceded ~46% of server-CPU revenue to AMD, and is a near-spectator in accelerators — its survival case rests entirely on whether the 18A node finds an anchor foundry customer. Placed beside NVIDIA’s $5.4T and Arm’s ~190× multiple, Intel at ~9× sales is the cross-section’s memento mori: leading-edge positions are not permanent, and the cost of losing one is the difference between owning the cycle and watching it.
Where they agree — and where they split
All six, even Intel’s bulls, accept that AI capex is real and large; the disagreement is entirely about durability and price. NVIDIA, Broadcom, AMD and Arm all trade at multiples (~20× to ~190× forward) that price in an uninterrupted super-cycle — so every one of their studies lands its valuation question as contested, with the same swing factor: the next capex turn, which none of them can yet observe. They split on who is most exposed when it comes: the bulls say diversified demand (Broadcom’s networking, Micron’s HBM, Arm’s royalty breadth) cushions it; the bears note that customer concentration is extreme at the top (NVIDIA’s ~61%) and that memory and merchant GPUs have always been cyclical. The demand for AI compute is not the question. Whether today’s prices survive the first air-pocket is.