Source Report

Nvidia's Early Capacity Lockup on TSMC's N3 Node

Dwarkesh Patel referenced SemiAnalysis modeling showing Nvidia capturing over 70% of TSMC's N3 (3nm) wafer capacity by 2027 through aggressive early commitments: Nvidia signaled massive AI demand via data center builds and signed non-cancellable contracts with deposits across the supply chain (e.g., PCBs, memory), securing priority as TSMC's highest-growth customer. This mechanism—prepaying for flex capacity expansions—displaced slower-moving clients, with TSMC prioritizing stable high-margin HPC/AI over mobile. SemiAnalysis' broader models confirm AI accelerators (Nvidia-led) consuming 86% of 2027 N3 output, squeezing smartphones/CPUs, though exact Nvidia share isn't broken out beyond Patel's projection.[1][2]
- SemiAnalysis projects Nvidia surpassing Apple's N3 consumption by 4Q27 as Apple shifts to N2 (Apple's N2 share falls to 48%).[3]
- TSMC 2024 Annual Report: N3 at 18% of wafer revenue (advanced nodes 69%), with HPC/AI driving expansions; top customer (likely Apple) at 22%, top-10 at 76%—no Nvidia specifics disclosed.[4]
For competitors entering N3 (e.g., hyperscalers), this means rationed allocations (5-10% uplifts for 2026), forcing GPU fallbacks or delays.

TSMC Customer Concentration and Public Wafer Disclosures

TSMC's disclosures reveal high concentration but no per-node breakdowns or Nvidia-specific N3 claims: top customer (Apple) at 22% total revenue (2024), top-10 at 76%, with North America (Nvidia-heavy) at 70%. No long-term supply agreements named; risks flagged include customer loss curtailing demand (e.g., via in-house shifts or export controls). N3 ramped to third-year volume in 2024 (N3E/P/X variants), with 17M total wafers; 2025 capex $38-42B targets N3/N2/CoWoS expansions (Taiwan/Arizona/Japan).[4]
- Analyst estimates (e.g., JPMorgan) align with Nvidia requesting 50% N3 uplift (100-110k to 160k wafers/month, ~35k for Blackwell/Rubin).[5]
Patel's 70%+ is model-derived (supply chain tracking), not public—plausible given Nvidia overtaking Apple as TSMC's largest overall customer (19% vs 17% revenue, 2025).[6]

Implication for Entrants: Without Nvidia-scale prepayments, new players face 2+ year lead times; diversify to Samsung/Intel or older nodes, but yields/pricing lag 20-30%.

Google TPU v7 (Ironwood) Production and Capacity Squeeze

Google's TPU v7 (Ironwood) on TSMC N3E ramps 2025-2026 (192GB HBM3E/chip, 4.6 PFLOPs FP8), but faces N3 tightness: SemiAnalysis notes "huge upsize" for 2026 (internal + external via Anthropic), yet reallocating just 5% smartphone N3 yields ~0.3M extra v7s—far short of demand. No public volumes; estimates ~1.35M v7 units in Google's 2026 TPU total (4.3M across variants, TSMC ~3.95M).[2][7]
- Matches/exceeds Blackwell per-chip specs (FLOPs/bandwidth close, pod-scale to 9k chips via ICI/OCS), but availability lags Nvidia by 1 year.[8]

Implication for Entrants: Google's internal TPU shortage forces GPU buys; scale requires TSMC lobbying or N2 shift, risking 25%+ cuts like reported CoWoS constraints.

Anthropic TPU Deal Timeline and "Selling Capacity" Narrative

Anthropic-Google deal announced Oct 23, 2025: up to 1M TPUs (~1GW online 2026, tens of $B value), expanding 2023 partnership. SemiAnalysis: ~400k v7 Ironwoods as $10B Broadcom racks + 600k GCP-rented; leverages ex-Googlers' info asymmetry for early access before Google's Gemini revenue exploded (Q4 2025: $5B ARR from $0).[9][1]
- Timeline: Negotiated pre-Google's demand inflection; no "sold before realizing own needs" evidence—deal mutual (Google gains revenue/backlog $49B QoQ jump).[4]

Implication for Entrants: Multi-cloud hedging (Anthropic mixes TPUs/Trainium/GPUs) works short-term, but frontier labs need 1GW+ commitments 18-24 months ahead.

Validity of Patel's Overall Narrative

Patel's story—Nvidia's early AGI-pilled bets locking 70%+ N3 by 2027, squeezing late Google (chip delays, conservative commitments)—holds mechanistically (prepayments/data signals), but 70% is SemiAnalysis projection (high confidence from tracking, aligns with 86% AI total). Google "squeezed" via N3 convergence (v7/Trainium3 2026), but ramps ahead (millions v7s); Anthropic deal post-dates Google's revenue surge, not "regret"—strategic diversification. TSMC disclosures confirm concentration/HPC boom, no contradictions. Speculative reconstruction fits (Google woke late per Patel), but oversimplifies: Google scales TPUs (Gemini 3 success), Nvidia leads via ecosystem/ramp speed.[1][2]

Implication for Entrants/Competitors: Commit 24+ months early with deposits; hedge multi-node/multi-vendor (N3/N2, GPU/ASIC); non-obvious: power > wafers now, but N3 famine forces N2 rush (Apple/Google fight). Confidence: High on squeeze mechanics/dominance; medium on exact 70% (modeled). Additional supply chain audits needed for 2027 precision.

---

Recent Findings Supplement (March 2026)

Nvidia's Dominance in TSMC N3 Capacity

SemiAnalysis models project Nvidia securing the majority of TSMC's N3 wafer capacity through early, aggressive long-term commitments and deposits, enabling Rubin GPUs (on customized N3P) to ramp massively in 2026-2027; this mechanism—prepaying for expansions while competitors like Apple grow slower—creates a de facto data moat, as TSMC prioritizes volume AI orders over stable but lower-scale smartphone chips, leaving non-AI demand (e.g., CPUs) displaced to N2 transitions.[1][2]
- AI accelerators (Nvidia Rubin, Google TPU v7/v8 on N3E, AMD MI350X/MI400, AWS Trainium3 on N3P) consume ~60% of N3 wafers in 2026 (rising to 86% in 2027), squeezing smartphones from 40% to near-zero; Nvidia drives the bulk via Rubin transition from 4NP Blackwell.[1]
- Dylan Patel (SemiAnalysis) states Nvidia holds "+70% of N3 wafer capacity by '27," based on supply chain tracking (e.g., PCB/memory vendors, non-cancelable contracts); no public TSMC confirmation, but aligns with Nvidia overtaking Apple as top overall customer (19% of 2025 revenue, $23.2B vs Apple's 17%).[2][3]
Implications for Competitors: New entrants must lock multi-year capacity now (2-3 year fab lead times), but Nvidia's scale (~200k-300k Rubin GPUs in 2026 despite constraints) blocks late movers; TSMC Q4'25 call warns of "very tight" N3 through 2027, prioritizing AI via productivity tweaks (e.g., N5-to-N3 conversions).[4]

Google TPU v7 (Ironwood) Faces Allocation Squeeze

Google's TPU v7 Ironwood (TSMC N3E, 192GB HBM3E, 4.6 PFLOPS FP8) ramps in 2026 amid full N3 convergence, but late internal demand inflection (Gemini revenue from ~$0 to $5B Q4'25) led to denied expansions—TSMC offered only 5-10% more for 2026, shifting focus to 2027; this forced Google to allocate ~1M units to Anthropic via prior ex-Google team ties, highlighting info asymmetry where labs preempt hyperscalers.[2][5]
- Ironwood production already up in 2025 (huge 2026 volumes via internal/external sales); SemiAnalysis notes Google "can't get them fabbed" at scale, with TPU v8 also N3-tied; Anthropic deal (Oct'25): 1M TPUs (~$10B direct racks via Broadcom + $42B GCP RPO), >1GW online 2026, but split eases Google's self-supply burden.[5]
- No "sell before realizing revenue explode" evidence; Anthropic's ex-Google ties secured pre-inflection, per Patel—Google requested Q3'25 hikes post-Gemini but got limited amid Nvidia/AMD priority.
Implications for Competitors: Google must diversify (e.g., neocloud backstops like Fluidstack/TeraWulf) or risk TPU shortages; labs like Anthropic gain leverage via multi-cloud (TPU + Trainium + Nvidia), but TSMC "kingmaker" role favors early AI volume over late hyperscaler pivots.[1]

TSMC's Customer Concentration Signals Shift

TSMC's 2025 annual report confirms Nvidia ("Client A") at 19% revenue ($23.2B, 2x YoY growth), overtaking Apple ("Client B") at 17% ($20.7B); top-10 now ~76% total, with AI/HPC at 58% wafer revenue (high-teens% from accelerators), but no N3-specific breakdowns—Q4'25 call notes N3 at 28% wafers, "sold out" through 2027 via CapEx pull-forwards ($52-56B'26, 70-80% advanced nodes).[4][3]
- No regulatory/policy changes; AI "multi-year megatrend" validated by customer data (cloud providers "making more money than TSMC").
- Arizona N3 volume stays 2028 (not relieving 2026-27 tightness).
Implications for Competitors: High concentration (Nvidia/Apple ~36%) amplifies risks, but TSMC's pricing power (N3 wafers $25-27k) rewards AI scale; late entrants face premiums or reallocations (e.g., smartphone-to-AI shifts yielding 0.1-0.7M extra Rubin GPUs).[1]

Assessment of Patel's Narrative

Patel's claim (Mar'26 podcast) has partial factual basis in SemiAnalysis models (Nvidia majority N3 via early locks; AI 86% by'27) and TSMC data (Nvidia top customer), but 70%+ is model-derived (supply chain signals), not disclosed—supported indirectly by N3 tightness and Rubin priority.[2] Google's "squeeze" aligns (late requests denied), with Anthropic TPU sale (Oct'25) fitting "sold capacity before realizing explosion" via ex-team foresight, not misjudgment; timeline: Gemini inflection Q4'25 post-deal, but no evidence Google foresaw/ignored—speculative reconstruction, as TSMC prioritizes regardless. Overall: Compelling but oversimplified; mechanism (early commitments) real, but Google ramps TPUs amid constraints, not "out."[5]
Implications for Competitors: Validates early-bird moat; compete via Samsung/Intel diversification or neoclouds, but TSMC's 70%+ foundry share locks AI path—additional research on SemiAnalysis models strengthens confidence (high now via cross-sources).

Recent Announcements Validate Tightness

• TSMC Q4'25 (Jan'26): N3 28% wafers; AI CAGR mid-50s% to'29; CapEx supercycle signals no relief til'28.[4]
• Anthropic-Google (Oct'25): 1M Ironwood TPUs, >1GW'26—largest external TPU deal, but power/DC bottlenecks persist.[5]
• TSMC 2025 Report (Feb/Mar'26): Nvidia 19% revenue—new stat confirming shift.[3] Implications for Competitors: No easing; enter via merchant alternatives (e.g., Broadcom XPUs), but hyperscaler ASICs fragment demand without displacing Nvidia's lead.