Ep. 006 - The AI Silicon Shortage Explained (AI Supply Chain & Fabs) | Sravan Kundojjala, Ivan Chiam, Jordan Nanos

Summary

• The AI build-out’s binding constraint has migrated again — from CoWoS packaging (2023) to data center power (2023-25) to raw wafer-fab capacity now. Ivan Chiam’s framing: “power is no longer the biggest constraint,” and demand is compounding via better models and agentic workloads — Anthropic added $6B of ARR in February alone, mainly on Claude Code; SemiAnalysis itself burns “a few thousand USD a day” in tokens; and on-demand Hopper pricing is rising on a chip “nearly two generations old.”
• AI goes from 9% of TSMC N3 wafer demand in 2025 to 60% in 2026 and ~85-90% in 2027, and the smartphone release valve doesn’t work. Major accelerators converge on N3 across 2025-27: TPU v7 was already shipping in 2025, while Rubin, AMD MI400, TPU v8 and Trainium ramp through 2026-27. Ivan’s modeling shows reallocating even 25% of 2026 smartphone N3 wafers yields only ~700K Rubin GPUs or 1.5M TPU v7s: “really isn’t enough to move the needle significantly.”
• NVIDIA overtook Apple as TSMC’s largest customer in 2025, and TSMC is acting as kingmaker on allocation. Sravan Kundojjala says TSMC “burned their hands” on crypto in 2018 and now underwrites demand durability. Jordan framed the squeeze as forcing some customers to bring their own memory allocation to prove they can use wafers; Ivan calls procurement “a new competitive moat.” The dynamic favors giants and leaves Qualcomm, MediaTek and Chinese low-end customers struggling to secure DRAM.
• HBM stays structurally tight until meaningful capacity lands in the second half of 2027. HBM consumes 3x the wafer capacity per bit of commodity DRAM, rising to 4x with HBM4/4E, while NVIDIA’s requests for ~11 Gbps pin speeds that “many memory vendors fail to hit” keep the market “very, very tight” — Sravan adds that compressed smartphone demand will return and “a perfect storm is brewing.”
• The H100 rental market inflected instead of deflating: SemiAnalysis expected -30% in 2026, but one-year prices bottomed around $1.70/hr in October and have since risen 15-20% plus probably another 10% in March. Memory probably added 5-10% to server cost — increases “well beyond that” are “very clearly demand oriented” — and Jordan heard that one NeoCloud renewed an expiring H100 contract for four more years, implying an eight-year GPU life priced through 2030.
• The top-calling signals to watch: whether ramping GB300 inference capacity loosens the 1-4-year contract market, ROI failing to pencil, and hearing “this time is different.” Sravan’s cyclical dissent is the sharpest edge in the episode — a lot of double ordering is going on “in memory as well as in N3,” and “this industry never learns this lesson… booms and busts will happen” — yet even he says Claude Code made 2026 his most productive year, “almost like 10X.”
• CPO’s first act is scale-out, not the scale-up trade the market positioned for — and the TAM question is whether it triples. NVIDIA’s Kyber compresses 144 GPUs into a 600kW rack precisely to stay in copper range, so CPO connects racks (NVL576, Feynman NVL1152, eight racks each); if CPO also goes inside the rack “that basically triples your TAM,” but Lumentum must ramp ultra-high-power lasers 20-30x off an unknown base and Dan cautions “a lot’s been priced in.”

Deep dive

1. The bottleneck has migrated to the fab floor

• Ivan’s periodization of the build-out: distinct bottlenecks in sequence — CoWoS advanced packaging in 2023, data center power and physical space from 2023 through 2025, and now “not enough wafer fab capacity. Power is no longer the biggest constraint” — front-end capacity is the binding constraint of this phase.
• The demand side, per Ivan: better models driving adoption plus agentic workflows like Claude Code. SemiAnalysis burns “a few thousand USD a day” in tokens, Anthropic added $6B of annual recurring revenue in February alone, mainly driven by Claude Code, and on-demand Hopper pricing is shooting up on a chip “nearly two generations old.”
• Dan’s ROI grounding for why major hyperscaler capex cuts are “definitely not” coming: Claude Code tasks costing $5-7 replace “three or four hours of analyst time.” Hyperscalers are “full steam ahead” and thinking about Vera Rubin.

2. TSMC is kingmaker — and NVIDIA just dethroned Apple

• Sravan’s regime change: from the first iPhone until the A100, smartphones drove TSMC’s leading edge; now NVIDIA overtook Apple as TSMC’s largest customer in 2025 and HPC is outgrowing handsets by a big margin. TSMC had been at almost $30B in CapEx before raising this year’s figure to almost $54-55B, with a stated range of $52-54B; SemiAnalysis expects more and likely ~$70B in 2027 — “will that be enough? I don’t think so.”
• On allocation politics: TSMC “burned their hands” on crypto customers in 2018 — demand vanished within two-three quarters — so it now studies whether a customer “has enough clout to drive demand for the next few years, not just one or two quarters.” Apple remains a predictable anchor at 10% of total wafer shipments and 25-30% of leading-edge demand; pricing is “value capture,” and Sravan insists TSMC is “not as opportunistic as people think.”
• Jordan framed the allocation squeeze as forcing some customers to bring their own memory allocation to prove they can use the wafers — good for giants, bad for smaller players. Ivan’s read is that procurement strategy is now “a new competitive moat.”

3. AI eats N3 — 9% to 60% to ~90% in two years, and no release valve

• The chart Jordan put on screen: AI was 9% of N3 wafer demand in 2025, 60% in 2026, and ~85-90% in 2027 — “this really squeezes out smartphone and consumer electronics.” Ivan’s mechanism is simultaneity across the ramp: Blackwell’s 4NP moves to Rubin on N3, AMD MI400 uses N3 for compute dies and AIDs, TPU v7 had already been shipping on N3 in 2025 with v8 to follow, and Trainium ramps from the second half of 2026.
• Sravan’s baseline: TSMC exited 2025 at 120K N3 wafer starts a month, roughly two-thirds consumed by smartphones and PCs — the mix that now flips toward accelerators.
• No near-term fix: capacity invested today takes a “minimum twelve to twenty-four months” to come online, and there is no immediate alternative foundry — Samsung is still struggling with 3nm yields. Best triage per Sravan: migrate Apple/Qualcomm premium tiers to N2, with TSMC trying to accommodate them at reasonable pricing to free N3, and keep mid-range on 4nm. Sravan said they had heard some game consoles were being pushed.
• Ivan’s modeling kills the smartphone-release-valve hope: reallocating 5% of 2026 smartphone N3 wafers yields only slightly over 100K Rubin GPUs or slightly over 300K TPU v7s; even an extreme 25% reallocation gets ~700K Rubins and 1.5M TPU v7s — “really isn’t enough to move the needle significantly.”

4. Downstream carnage in handsets; HBM tight into 2H27

• Sravan on the consumer side: memory has climbed from 17-20% to 25-30% of handset bill of materials; Xiaomi, OPPO and Vivo are cutting low-end orders by up to 30%, with smartphone and PC units probably down 10-15% this year. “Flagship is safe… the low end is very much taking a hit” — it is whoever lacks volume and supply-chain leverage.
• Ivan’s house view on memory: meaningful capacity only arrives in the second half of 2027. The math: HBM consumes 3x more wafer capacity per bit than commodity DRAM, rising to 4x with HBM4/4E, while HBM content per chip climbs every generation.
• The tightening flywheel: NVIDIA requests very high pin speeds — e.g. 11 Gbps — and “many memory vendors fail to hit that requirement,” keeping HBM “very, very tight… a structural trend.”
• Sravan’s twist: compressed smartphone demand is ironically relieving wafer pressure now — “otherwise this would have been even worse” — but that demand comes back: “a perfect storm is brewing,” with no relief absent “huge capex cuts by hyperscalers, which is unlikely.”

5. The rental market inflected — GPUs now priced through 2030

• Dan expected H100 rentals to fall 30% in 2026 as GB300s lowered cost per compute; instead one-year prices bottomed around $1.70/hr/GPU in October, ticked to ~$1.80, then “rocket[ed] up” 15-20% in the first couple of months plus probably another 10% in March.
• Sravan’s ISO-IRR estimate is that memory probably added 5-10% to server cost, so increases well beyond that are “very clearly demand oriented.”
• Jordan’s sourcing experience: like “trying to find airplane tickets on the last flight out” — based on what NeoClouds were saying, no H100s were coming off contract and no uncommitted capacity was coming online until roughly August-September. Jordan heard that one NeoCloud renewed an expiring H100 deal for a four-year term — GPUs signed through 2030, an eight-year life that rewrites the financial modeling and terminal-value debate.
• Sravan’s three-tier map: 4-5-year offtake contracts signed by large AI labs such as OpenAI and Anthropic, typically for GB300s and hundreds of megawatts; a 1-year to 3- or 4-year contract market, mostly AI natives but also some AI labs; and a thin on-demand tail. CoreWeave had reportedly renewed a lot of H100s late last year, keeping supply out of the market. The key test is whether ramping GB300 inference capacity satisfies demand and loosens that middle tier.

6. Three ways to call the top — and one honest cyclical dissent

• Dan’s tells: the four words “this time is different”; rental prices evolving more weakly than the compute-cost curve implies; and ROI failing to pencil. His bull kicker: “we’re really the tip of the spear” — Fortune 500 agent penetration is “extremely low,” and many enterprises lack even IT clearance to start.
• Ivan watches end demand — Anthropic/OpenAI user and revenue growth — and data center leases, citing Microsoft pulling out of non-binding contracts in 2025 as an early indicator. Verdict: “still very early in the cycle and still very bullish.”
• Sravan’s cyclical warning, worth keeping: a lot of double ordering is going on “in memory as well as in N3 capacity” — panicked customers order the same demand directly and through distributors, then cancel once served, leaving excess inventory for the other suppliers. “This industry never learns this lesson… booms and busts will happen.”
• Even while stressing cyclicality, Sravan says Claude Code has made 2026 his best year, “almost like 10X… it’s like I’m always having two, three people working for me” — dashboards from 40-50-tab Excels are now “child’s play.”

7. CPO arrives scale-out first — the market had the trade backwards

• Dan’s explainer: co-packaged optics moves the optical engine onto the substrate next to the chip, shortening the electrical path and allowing the system to avoid DSPs and the pluggable faceplate — saving power, eventually cost, and enabling more bandwidth to escape. Copper’s reach at 224G is only ~2 meters, limiting high-bandwidth scale-up largely to a single rack.
• The surprise from GTC/OFC week: supply-chain watchers seeing optical engines being built assumed scale-up (Kyber) would lead; instead NVIDIA’s vanguard is scale-out CPO, marketed to NeoClouds that already buy end-to-end NVIDIA solutions — multiplane CPO switches reaching 409 Tb aggregate versus ~100 Tb for a Tomahawk 6-generation switch.
• The Kyber oxymoron, in Dan’s words: “you spend all this effort to compress it into 600 kilowatts… get everything within copper range, and then you don’t use copper?” So optics connects racks: the NVL576 (Rubin, eight racks using scale-up CPO) and Feynman’s NVL1152 (eight Kyber racks).
• The TAM debate investors hung on: is Feynman’s CPO between racks only, or everywhere including between GPUs and NVLink switches? The latter “basically triples your TAM.” Lumentum must ramp ultra-high-power lasers 20-30x — but “no one knows what the base is,” and after the share-price runs “a lot’s been priced in.” Jordan’s sociology: investors hang on Jensen’s every word while pragmatic engineers see CPO as inevitable but do not know whether it lands in 2027 or 2028.

8. Dueling MSAs — but “all three will coexist”

• The OCI MSA — “all the right logos”: NVIDIA, Broadcom, AMD, Meta, Microsoft, OpenAI — signals a pivot from DR optics, described as “kind of gray optics,” to DWDM: Sravan tentatively read the displayed specification as four transmit and four receive wavelengths, bidirectional at 50G NRZ, fitting more traffic per fiber. Dan’s “final boss” for OCI is a die-to-die connection at NRZ, “keep it slow and wide” — “but we don’t think we’re quite there yet.”
• The CPX MSA defines the physical connector and form factor while aiming to stay silent on modulation — though Dan flags the curiosity that it “seems to presume ring modulators.” XPO is Arista-led, extending pluggables with strong cooling and density, possibly enabling linear pluggable optics by Tomahawk 7 — “a few years out.”
• Dan’s close, answering Jordan’s request for industry drama: “they’ll all coexist… they’re just different approaches, different ways to play the game.”