Nvidia vs Broadcom: the title fight for AI silicon

Every reported figure below is dated to its fiscal quarter. Nvidia's fiscal year ends in late January — its latest reported quarter is fiscal Q2 2027, ended July 26, 2026, reported August 26, 2026. Broadcom's fiscal year ends around late October — its latest reported quarter is fiscal Q3 2026, ended August 2, 2026, reported September 2, 2026. The two quarters end one week apart, so the latest-quarter comparison is genuinely like-for-like in calendar time. Anything for a later quarter is labelled guidance or estimate, and is never mixed into a reported comparison without saying so.

The verdict — the pairing in numbers

This is not a close fight on scale, and it is a close fight on almost everything else.

Where the fight is NOT close: absolute AI revenue. Broadcom's AI line is 18.8% of Nvidia's Data Center line in the latest quarter (16.7 ÷ 89.023) and only 15.3% on a trailing-twelve-month basis (42.4 ÷ 277.80). The one-quarter AI gap alone — $72.3 billion — is larger than Broadcom's entire fiscal 2025 revenue ($63.9 billion).

Where the fight IS close, or Broadcom leads: profitability per dollar (non-GAAP margins are level or better for Broadcom; its GAAP margins look worse only because of ~$2.0 billion a quarter of acquisition-intangible amortization from the VMware deal), cash conversion, growth rate (+221% vs +117% on the AI lines), and — remarkably — valuation per sales dollar: the market pays roughly the same price-to-sales for Broadcom as for Nvidia (see Section 06), so Broadcom's smaller market value is a size discount, not a quality discount.

The direction of travel is the real story. On TTM figures Broadcom's AI line is 6.55× smaller; on the latest quarter it is 5.33× smaller; on the two companies' own next-quarter guidance (labelled as such: Broadcom AI $21.7B guided for Q4 FY26 vs Nvidia Data Center implied by its $108.0B total guide) it would narrow further. Broadcom management guides its AI line to ~$58B in fiscal 2026, ~$115B in fiscal 2027 and ~$230B in fiscal 2028 — company claims, not results (Section 04). Even if every one of those numbers lands, Broadcom's AI line in fiscal 2028 would still be below Nvidia's Data Center TTM today ($277.8B) — and Nvidia guides its own total revenue to grow ~70% in fiscal 2028. Closing the gap therefore requires Broadcom to hit a two-year doubling path and Nvidia's growth to fall well below its own guidance at the same time.

Section 01The two models — what each company actually sells into AI

NVIDIA: merchant GPU + rack system + NVLink + CUDA

Nvidia sells the same product to everyone: general-purpose GPUs (Hopper, then Blackwell/Blackwell Ultra, now Vera Rubin), assembled increasingly into rack-scale systems (GB200/GB300 NVL72, Vera Rubin NVL72) in which the GPU, the Vera CPU, NVLink/NVSwitch scale-up fabric, Spectrum-X Ethernet and InfiniBand scale-out networking, and the CUDA software stack are one product. The Q2 FY27 release describes Vera Rubin as "now in full production" with "racks running at partners including CoreWeave, Google Cloud, Microsoft Azure, Oracle Cloud Infrastructure and Nebius," and lists Spectrum-6 switch systems and the Groq 3 LPX inference accelerator as part of the same platform (Nvidia Q2 FY27 release, Aug 26, 2026).

Revenue mix, latest reported quarter (Q2 FY27):

• By market platform: Data Center $89.023B = 92.5% of revenue; Edge Computing (gaming, professional visualisation, automotive, robotics) $7.198B = 7.5% (Nvidia release; segment recast from Q1 FY27).
• By reportable segment: Compute & Networking $88.299B (91.8%), Graphics $7.922B (8.2%) (Nvidia CFO commentary, via TickerGrove's filing summary and note.com analysis of the 10-Q).
• Within Data Center, by customer class (new disclosure): Hyperscale $48.710B (54.7% of Data Center), AI Cloud/Industrial/Enterprise ("ACIE") $40.313B (45.3%) (same sources).
• Networking is now a giant business inside Data Center: fiscal 2026 Data Center revenue of $193.7B split into compute $162.4B and networking $31.4B, up 142% (Nvidia fiscal 2026 10-K figures, via FourWeekMBA and Motley Fool). Secondary reporting of the Q2 FY27 call put Q2 networking at ~$17.5B (Q1's $14.8B +18% sequentially) — that figure is an estimate derived from the call, because Nvidia stopped breaking networking out in Q2; it is labelled as such wherever it appears.

The software layer (CUDA, CUDA-X libraries, NIM/Agent Toolkit, cuDNN) is not broken out as revenue — it is sold through the hardware margin (75.0% gross margin). That is the platform model: Nvidia captures the full system value and the developer ecosystem, and the customer gets a chip that can run any workload and be resold or rented to anyone.

Broadcom: custom XPU design + merchant networking silicon + VMware

Broadcom sells three different things into (and around) AI:

1. Custom accelerators ("XPUs"). Broadcom co-designs an ASIC to one customer's workload — Google's TPU, Meta's MTIA, OpenAI's "Jalapeño" — supplying the silicon design, SerDes, packaging and networking IP; the customer owns the architecture and the chip is useless to anyone else. Broadcom confirmed six XPU customers on its Q2 FY26 call (June 3, 2026); four are publicly named — Google, Meta, OpenAI, Anthropic — and two remain unnamed, with analysts most frequently identifying Apple and ByteDance as the candidates (Axis Intelligence summary of the Q2 call; Teardown customer table). In Q3 FY26, XPUs were 73% of AI semiconductor revenue (≈$12.2B) with AI networking the other ≈27% (≈$4.5B) (Zacks summary of Q3 results, Sep 3, 2026; NextPlatform split of $12.22B compute / $4.52B networking cited by Tech-Insider).
2. Merchant networking silicon — the Tomahawk/Jericho class. Broadcom sells Ethernet switch and routing silicon to everyone, including customers who build on Nvidia GPUs: Tomahawk 6 (102.4 Tb/s, 3nm, production shipments from March 2026), Tomahawk Ultra (51.2 Tb/s scale-up Ethernet), Jericho4 (deep-buffered scale-out routing) and TH6-Davisson co-packaged optics (Broadcom announcements via IEEE ComSoc Tech Blog, Oct 2026, and TechTimes Q3 coverage). This is the "open Ethernet" counter-position to Nvidia's proprietary NVLink/InfiniBand fabric — and it means Broadcom earns AI networking revenue even from clusters whose compute is Nvidia's.
3. Infrastructure software — the segment Nvidia has no equivalent of. Built around VMware (acquired for $61.4B in November 2023), plus CA and Symantec franchises: VMware Cloud Foundation private-cloud subscriptions. Nvidia sells software only as part of its hardware platform; Broadcom has a standalone, recurring-revenue software business larger than most software companies.

Revenue mix, latest reported quarter (Q3 FY26, primary release):

• Semiconductor solutions $20.839B = 70% of revenue (+127% y/y); Infrastructure software $8.752B = 30% (+29% y/y).
• AI semiconductor revenue $16.7B = 56.4% of total revenue (16.7 ÷ 29.591) and 80.1% of semiconductor revenue; non-AI semiconductor revenue was just $4.2B (+5% y/y, flat sequentially — broadband, server storage, wireless; Broadcom Q3 call via SemiBoard transcript).
• Software quality: infrastructure software ran at an ~84% segment operating margin in Q3, with ARR growth of 15% y/y (Q3 call, via SemiBoard/InvestReport summaries). For fiscal 2025 as a whole, software was $27.0B of the $63.9B total (Broadcom Q4 FY25 call).

The structural difference in one line: Nvidia's model concentrates 92.5% of revenue in one platform it fully owns; Broadcom's model spreads revenue across a custom-silicon design business (a handful of customers), a merchant networking franchise (everyone), and a software annuity (no AI exposure at all) — $8.8B a quarter that Nvidia has no answer to and that funds Broadcom's cash flow through any AI capex pause.

Section 02The scale and profit comparison

Latest reported quarters, side by side — both ended within a week of each other.

Three caveats that change how the table should be read:

1. The GAAP margin gap is an accounting artefact, mostly. Broadcom's GAAP gross margin is depressed by $1,499M of acquisition-intangible amortization (VMware) plus $224M stock compensation in cost of revenue in the quarter alone; on the non-GAAP basis both companies use for guidance, gross margins are identical at 75.0% and Broadcom's operating margin is higher. Conversely, Nvidia's GAAP net income ($59.688B) is above its non-GAAP net income ($53.954B) this quarter because GAAP includes $7.771B of gains on equity securities — paper gains on its investment book. Neither GAAP bottom line is the clean one this quarter.
2. Nvidia's Q2 free cash flow is the misleading number in the table. Operating cash flow fell to $24.077B in Q2 (from $50.344B in Q1) because accounts receivable jumped $22.346B in the quarter as revenue grew 18% sequentially; Q1 FCF was $48.554B (59.5% of revenue). The TTM comparison below is the fair one for cash flow.
3. Guidance is not in this table. Nvidia guides Q3 FY27 revenue to $108.0B ±2% at 74.0% gross margin, assuming no China Data Center compute revenue; Broadcom guides Q4 FY26 revenue to ~$34.8B (+93%) with AI revenue of $21.7B (+236%) and non-GAAP operating margin ~66% (both from the primary releases). On guidance alone the total-revenue ratio would narrow from 3.25× to 108.0 ÷ 34.8 = 3.10× — but that is a guided-vs-guided comparison, not a reported one.

The like-for-like AI-silicon comparison — and which comparison is fairer

Broadcom AI semiconductor revenue vs Nvidia Data Center revenue is the fairer fight, and it is also the one Broadcom loses by more. It is fairer because it compares the two AI infrastructure businesses rather than two whole companies: Nvidia's total includes $7.2B a quarter of Edge Computing and Broadcom's total includes $8.8B a quarter of VMware software and $4.2B of non-AI chips, none of which compete with each other. On that basis: latest quarter 5.33×, TTM 6.55×, Broadcom at 18.8% / 15.3% of Nvidia's line.

Each point pairs the two companies' reported quarters by recency (the fiscal calendars differ; the latest pair ended one week apart). The ratio is computed from the reported AI lines: $51.215B ÷ $6.5B, $62.314B ÷ $8.4B, $75.246B ÷ $10.8B, $89.023B ÷ $16.7B. The direction is Broadcom's; the level is still Nvidia's.

Two imperfections in even that comparison, both of which flatter Broadcom if ignored:

• Broadcom's "AI semiconductor revenue" includes its AI networking silicon (~27% of the line), some of which is sold into clusters built on Nvidia GPUs — revenue that is complementary to Nvidia, not taken from it. Nvidia's Data Center line likewise includes $31.4B (FY26) of networking. The pure compute-vs-compute comparison (Broadcom XPU ≈ $12.2B vs Nvidia Data Center compute, ~$71.5B if Q1's compute/networking split is extrapolated — an estimate, not a reported Q2 split) would be wider still, roughly 5.9×.
• The total-company comparison (3.25× / 3.40×) is the right one for enterprise value and cash generation — because VMware's $8.8B a quarter at ~84% operating margin is real money that funds Broadcom's AI R&D and debt paydown — but the wrong one for "who is winning AI silicon."

Plain answer: for the title fight, use the AI-line comparison (Broadcom is one-fifth to one-sixth Nvidia's size and closing); for solvency, cash flow and valuation, use the total-company comparison (Broadcom is roughly one-third Nvidia's size, with equal or better margins per dollar).

Section 03The customer overlap — where they actually collide

Broadcom has six confirmed XPU customers; four are named (Google, Meta, OpenAI, Anthropic) and two are not (reported candidates: Apple, ByteDance). Nvidia does not name customers in its releases, but named deployments, partner announcements and the customers' own disclosures identify the same names. Status labels: Signed = announced by a company or in a filing; Reported = press or analyst reporting, not company-confirmed.

Google / Alphabet Signed

Buys from Broadcom: TPU co-design since ~2014; April 6, 2026 8-K: long-term agreement to develop and supply future TPUs (v7 Ironwood, v8t/v8i) and networking/rack components through up to 2031. Ironwood delivered "in high volume" in 2026 (Broadcom Sep 2026 call). This is Broadcom's anchor XPU account. Buys from Nvidia: Google Cloud buys Nvidia GPUs to rent — Blackwell instances live; Oct 21, 2026 Google Cloud/Nvidia announcement of Vera Rubin NVL72 "A5X" instances (GA November 2026), with a plan for up to 960,000 Rubin GPUs. Gemini itself trains/serves mainly on TPUs. Collision? Partial. Direct collision for Google's own model compute (TPU wins); no collision for Google Cloud's rental fleet, where customer demand for CUDA keeps Nvidia in. Google is simultaneously Broadcom's biggest design win and a top-tier Nvidia cloud customer.

Meta Signed

Buys from Broadcom: MTIA co-design; April 14, 2026 extension through 2029, initial commitment >1 GW, "first phase of a sustained multi-gigawatt rollout," first 2nm custom AI chips; MTIA 300 in production for ranking/recommendation, 400/450/500 to follow (Meta/Broadcom announcements). Broadcom Q2 call framed Meta at ~3 GW of XPUs through end-2028. Buys from Nvidia: Feb 17, 2026 multiyear deal for millions of Blackwell and Rubin GPUs plus Spectrum-X networking (Nvidia newsroom). Collision? Yes — the cleanest head-on account. Meta publicly runs a three-vendor strategy (Broadcom MTIA + Nvidia + up to 6 GW of AMD). Workloads are partitioned: MTIA takes ranking/recommendation and growing inference; frontier training stays on Nvidia. Every MTIA gigawatt is a gigawatt Nvidia bid for and didn't get.

Anthropic Signed

Buys from Broadcom: Revealed Dec 2025 as Broadcom's "$10B customer." Oct 2025: up to 1M Google TPUs, "tens of billions," >1 GW online 2026. April 6, 2026: expanded agreement with Google + Broadcom for ~3.5 GW of next-gen TPU capacity from 2027 (Broadcom 8-K; consumption conditional on Anthropic's continued commercial success). Broadcom Sep 2026 call: 1 GW Ironwood deploying in 2026, expects Anthropic to become its largest XPU customer in 2027. Anthropic's S-1 (filed Oct 26, 2026) lists a ~$161.2B lease for AI accelerator equipment with Broadcom among its compute contracts (via Seoul Economic Daily and Beating The Tide's S-1 read). Mizuho estimate: Anthropic alone could be $21B of Broadcom AI revenue in 2026, $42B in 2027 — analyst estimate, not guidance Estimate. Buys from Nvidia: Nov 18, 2025: Anthropic commits $30B to Azure compute and contracts up to 1 GW of Grace Blackwell / Vera Rubin capacity; Nvidia invests up to $10B in Anthropic. S-1 read: Microsoft sells Anthropic Nvidia capacity on Azure for at least $44.9B; a separate ~$45B Nscale Vera Rubin deal (reported). Anthropic also runs on AWS Trainium (>$100B/10yr AWS agreement reported from S-1). Collision? Yes, but multi-sourced by design. Anthropic deliberately splits across TPU (Broadcom/Google), Trainium (Amazon/Marvell) and Nvidia (via Azure/neoclouds). Broadcom is its largest single silicon channel on the S-1 numbers; Nvidia capacity is real but bought through clouds, not direct.

OpenAI Signed — term sheet

Buys from Broadcom: Oct 13, 2025: co-develop and deploy 10 GW of OpenAI-designed custom accelerators + Broadcom Ethernet networking, racks targeted H2 2026–end 2029; OpenAI designs, Broadcom develops/manufactures. First fruit: "Jalapeño" inference chip, unveiled June 2026, design-to-tape-out in nine months (OpenAI claim), volume expected 2027. Financing is the soft spot: Oct 2026 reports (WSJ/Bloomberg, via TechArcade) have Broadcom working to arrange $30–50B+ of debt so OpenAI can pay for the chips — reported talks, nothing signed Reported. Buys from Nvidia: Sep 22, 2025 LOI: ≥10 GW of Nvidia systems, Nvidia to invest up to $100B progressively; still an LOI (no definitive agreement) as of Dec 2025 per CFO Kress. Feb 2026: Nvidia instead took a $30B equity stake in OpenAI's funding round (reported, VLSI/Tech-Insider). Aug 2026 10-Q: Nvidia guarantees up to $105B tied to an OpenAI-linked Ohio campus (reported from filing). Collision? Yes — the largest contested account in AI. Both sides claim a "10 GW" OpenAI future; neither is a completed deployment. Broadcom's is a supply term sheet with financing unresolved; Nvidia's is an LOI partially converted into equity. OpenAI also signed 6 GW with AMD — it is buying optionality from everyone.

Microsoft — no confirmed Broadcom XPU relationship

Buys from Broadcom: No confirmed XPU relationship. Microsoft's Maia accelerators are widely reported as a Marvell design partnership, not Broadcom. Broadcom sells Microsoft networking silicon and VMware software as it does other enterprises, but no named custom-AI-silicon deal was found in any source. (One Motley Fool article loosely lists Microsoft among companies Broadcom is "working with" on custom silicon — uncorroborated; treated here as Rumored / unsupported.) Buys from Nvidia: Azure is a launch partner for Vera Rubin racks (named in Nvidia's Q2 FY27 release); Microsoft is one of the hyperscalers inside Nvidia's $48.7B-a-quarter Hyperscale line and the channel for Anthropic's 1 GW Nvidia commitment. Maia 200 is deployed for Copilot/OpenAI inference; Maia 300 (2nm, production 2027) was unveiled Oct 12, 2026 — via TSMC/Marvell ecosystem, not Broadcom. Collision? Mostly no collision with Broadcom. Microsoft is where the custom-silicon threat to Nvidia runs through Marvell, not Broadcom. Including Microsoft in "Broadcom's customer list" would overstate Broadcom's roster.

Apple Signed — non-AI Reported — AI

Buys from Broadcom: July 2026: Apple multiyear commitment >$30B through 2031, >15B US-made chips, ~$1.5B Fort Collins expansion — but the announced scope is custom silicon components and wireless connectivity (RF/connectivity), i.e. the legacy Apple business. Separately Reported: Apple is developing its first AI server chip ("Baltra") with Broadcom, 3nm, mass production H2 2026 for 2027 deployment (press/analyst reporting since Dec 2024, not Apple-confirmed). Apple is also the leading candidate for one of Broadcom's two unnamed XPU slots. Buys from Nvidia: Nvidia's Q2 FY27 release: "NVIDIA GPUs with Confidential Computing are now used for confidential inference in Apple's Private Cloud Compute." No large Apple GPU purchase has been reported. Collision? Not yet. Today Apple pays Broadcom for phone silicon and Nvidia for a niche confidential-computing deployment. It becomes a collision only if the reported Baltra ASIC displaces future Nvidia capacity in Apple data centres — a 2027 story, currently unconfirmed by Apple.

ByteDance Reported

Buys from Broadcom: Named by press/analysts (from Dec 2024) as one of Broadcom's original three hyperscaler XPU customers, co-developing two custom chips with Broadcom/TSMC for TikTok/Douyin inference, expected 2026. Broadcom has never named it. Status in 2026 is muddied: Oct 15, 2026 Bloomberg reporting has ByteDance in advanced talks with MediaTek for custom ASICs (unsigned, unconfirmed) after a reported May 2026 Qualcomm ASIC deal — suggesting diversification away from, or beyond, Broadcom. Buys from Nvidia: Reported 2026 Nvidia budget of ~¥100B (~$14.3B), focused on H200 under the January 2026 US licensing regime (case-by-case licences, 25% duty, per-firm caps); ~10,000 H200s reportedly delivered (to Hong Kong) by August 2026. Nvidia books ~zero China Data Center compute revenue in its outlook, so little of this is in Nvidia's reported numbers. Collision? A collision in a market neither can fully serve. Export controls, not product merit, decide this account (Section 06). Both relationships are press-reported rather than company-confirmed.

Amazon / AWS (context)

No XPU relationship found — Trainium/Inferentia are Marvell-partnered. Broadcom networking/software only. AWS is a major Nvidia buyer (AWS offers Blackwell/Rubin; reported 2M additional Nvidia GPUs with Vera CPUs deploying from 2026). No Broadcom collision; the second proof (with Microsoft) that "custom ASIC" ≠ "Broadcom."

xAI / SpaceXAI Reported

Buys from Broadcom: Press/analyst reports of a major Broadcom ASIC order for a custom inference chip, mass production 2027 — unconfirmed. Buys from Nvidia: Nvidia Q2 release: SpaceXAI will deploy Nvidia Vera CPUs; xAI's Colossus is an Nvidia GPU cluster, and Anthropic's S-1 reportedly includes up to $120.5B renting Colossus capacity. Collision? Potential future collision; today xAI is an Nvidia account.

Answer to the section's question: the two actually collide head-on in exactly three accounts — Meta, Anthropic and OpenAI — plus Google's internal-compute budget. Everywhere else they sell to the same customer without colliding: Google Cloud's rental fleet (Nvidia) vs Google's own TPU (Broadcom); Apple's phone RF chips (Broadcom) vs Apple's Private Cloud Compute GPUs (Nvidia); and the entire networking layer, where Broadcom Tomahawk silicon sits inside AI factories regardless of whose accelerator is in the rack — including, sometimes, Nvidia's own competitors' racks and white-box builds for Nvidia-GPU clusters.

Section 04The share contest — estimates, named and disagreed

There is no audited AI-accelerator market-share table. Revenue share flatters Nvidia (premium pricing; a TPU is an internal cost line, not a market sale), shipment share flatters ASICs (cheaper units). Every figure below is attributed, and the disagreements are the finding.

Shipment-based forecasts (Goldman, JPMorgan, TrendForce) converge on ASICs reaching roughly half of units by 2027; revenue-based estimates still give Nvidia three-quarters or more of dollars in 2026. Both can be true at once, because an ASIC unit books a fraction of a GPU unit's revenue (see Section 05) — and because Broadcom books only the design/silicon value of a TPU, while Google captures the system value internally.

Who wins the custom-ASIC business itself — estimates conflict here too

The 60% vs 80–85% gap is a denominator difference (all ASIC design services vs high-end/frontier ASICs), not a factual dispute about Broadcom's lead — every source puts Broadcom first by a wide margin.

The companies' own claims — attributed, not fact

Broadcom management claims:

• December 2024: serviceable addressable market (SAM) of $60–90B in fiscal 2027 from just three hyperscaler customers, each deploying 1M-XPU clusters (Hock Tan, Q1 FY25 call; reiterated through Q1 FY26).
• December 2025: $73B AI backlog over 18 months (Q4 FY25 call).
• June 2026: AI revenue ~$56B guided for FY26, >$100B "line of sight" for FY27; >$30B of AI bookings in Q2 alone (Q2 FY26 call).
• September 2, 2026 (latest, Q3 call/release): FY26 AI guidance raised to $58B; ~$115B for FY27 and ~$230B for FY28, with Tan stating supply is secured for both and demand exceeds the FY27 figure; non-GAAP EPS >$30 "on track" for FY28 (via AInvest/Zacks/Teardown summaries of the call). These are management statements on an earnings call — the FY28 figure is a two-year-out claim with no prior Broadcom precedent for guiding that far.

Nvidia management claims:

• GTC, March 16, 2026: "at least $1 trillion" of Blackwell + Vera Rubin purchase orders / demand visibility through 2027, up from $500B claimed at GTC 2025; CFO Kress confirmed the figure covers only Blackwell + Rubin plus related networking — not Feynman, Groq or standalone CPUs (Reuters and DataCenterDynamics coverage of the keynote; Constellation Research analyst-meeting notes quoting Huang: "strong visibility of $1 trillion plus of Blackwell plus Rubin").
• August 26, 2026 call: fiscal 2028 revenue growth of ~70%, explicitly described as supply-constrained ("our entire supply chain is challenged"), Vera Rubin the fastest ramp in company history at ~20% of Q3 FY27 Data Center revenue, and revenue per gigawatt rising $18B (Hopper) → $25B (Blackwell) → $40B (Vera Rubin) (call reporting via BetaFinch, InvestReport, NyxtBig).
• Longer horizon: AI infrastructure spending of $3–4 trillion by 2030 (Huang, reiterated Oct 2026).

Neither company's visibility claim is a backlog disclosure in a filing; both should be read as management framing. The honest cross-check is the reported run-rate: Nvidia's claimed $1T through 2027 against $277.8B of TTM Data Center revenue is plausible arithmetic if growth holds; Broadcom's claimed $230B FY28 AI revenue against a $42.4B TTM AI base requires the line to grow 5.4× in two years — faster than it grew in the last two ($12.2B FY24 → $20B FY25 → $58B guided FY26 is, however, a comparable slope).

Section 05Why customers go custom — and the counter-case

The economics (sourced)

• Unit cost. JPMorgan estimates (via press summaries) put Google's Ironwood TPU at ~$13,000 per chip for FP8 performance close to Nvidia's B200/B300, vs ~$30,000 for a high-end Nvidia accelerator; cloud providers' in-house chip purchase costs are estimated at ~40% of Nvidia/AMD GPU retail prices, with rental costs at 60–70% of GPU rentals. Analyst estimate, not a price list — neither Google nor Nvidia publishes these prices. Estimate
• Cost per token. SemiAnalysis's September 2026 InferenceX work (the first third-party TPU inference benchmark, via ForYou/Memeburn coverage): at 100 tokens/second per user, Ironwood served at $0.181 per million tokens vs $0.222 (B200) and $0.276 (B300) — 18–19% and ~34% cheaper respectively. The same work shows the limits: at a ~20–30-second median response-time target the gap narrows to ~8% or reverses (B200 ahead); Ironwood lacks native FP4; time-to-first-token at 256 concurrent users was 2.25× slower than B300. SemiAnalysis also estimates Google's internal TPU cost at ~$1.03/chip-hour vs a $12.00 public on-demand list price — i.e. the ASIC cost advantage is largest for the owner-operator and mostly disappears at cloud list prices. Estimate
• Vendor-stated TCO claims. Broadcom-reported figures circulating in analyst summaries claim custom ASICs deliver 30–50% lower TCO than Nvidia GPUs for specific workloads — a vendor claim, unattributed to an independent benchmark. Amazon makes the analogous claim for its own (Marvell-partnered) Trainium: 30–40% better price-performance per generation — also a vendor claim, explicitly noted as not independently confirmed in the coverage.
• The margin-transfer argument. Nvidia earns a 75% gross margin on the systems it sells. A hyperscaler designing its own chip pays TSMC, Broadcom and the HBM vendors instead — capturing a large part of that 75% itself. Analyst framing (Nodvolt, June 2026): the unit economics of designing your own chip become favourable above roughly 300,000–500,000 units a year — a threshold every hyperscaler has now crossed. This is an analyst estimate of the break-even, not a disclosed figure. Estimate

Supply control and strategy (sourced)

• Allocation independence. Custom programs let hyperscalers "partially route around Nvidia's allocation queue" (ValueAddVC ranking, citing Counterpoint/Silicon Analysts foundry-allocation reporting) and secure TSMC CoWoS/HBM capacity on their own contracts — Broadcom's July 2026 Samsung MOU (>$200B over five years for HBM, 2nm foundry and advanced packaging — an MOU, not booked orders) is explicitly a supply-security move for its XPU programs, diversifying beyond TSMC/SK hynix.
• Workload fit. ASICs are built for one stable, enormous workload — Meta's ranking/recommendation, Google's Gemini serving, OpenAI's inference — where generality is wasted silicon. OpenAI says Jalapeño was co-designed around its models, kernels and serving systems (OpenAI announcement, June 2026).
• Negotiating leverage. Even a credible ASIC program that never takes majority share disciplines Nvidia pricing and allocation. Meta's and Google's behaviour — expanding ASICs while signing record Nvidia deals in the same quarter — is consistent with leverage-seeking as much as replacement.

The counter-case (sourced)

• Flexibility and resale of capacity. A GPU runs any model, any framework, and can be rented to any customer — which is why Google Cloud, which owns the world's most successful ASIC, still buys Nvidia at scale "because many customers want it." A TPU or MTIA has one user and no resale market. Clouds monetise flexibility; internal ASIC fleets monetise only their owner's workload.
• Time-to-market and execution risk. Building a competitive accelerator from scratch takes roughly 3–5 years and >$1B before production (DC Atlas analysis), plus a software stack — compilers, kernels, framework integration — that must be maintained forever. Jalapeño's nine-month design-to-tape-out (OpenAI's claim) is the exception that proves the rule: it is inference-only, built on Broadcom's mature XPU platform, and was still only at engineering-sample stage in mid-2026 with final performance "still being measured" (OpenAI's own caveat).
• CUDA. The consensus moat description across analyst sources (Gyre Research competitive landscape, June 2026; DC Atlas): CUDA is "a software moat, not a silicon moat" — a decade-plus of developer lock-in, with most AI research and production code written CUDA-first. Nvidia's counter-moves extend the moat around ASICs rather than only against them: NVLink Fusion (March 2026) licenses NVLink into third-party custom chips (Marvell is a partner, with a reported $2B Nvidia investment), so even rival ASICs can end up paying Nvidia for interconnect, CPUs and NICs.
• Moving target. Vera Rubin's claimed economics — up to 10× lower inference cost per token than Blackwell on MoE workloads (Nvidia product claim, CES January 2026; repeated in Q2 coverage as 35× token-cost reduction vs Blackwell Ultra for the full platform generation — vendor figures, workload-specific) — mean an ASIC designed against Blackwell's economics in 2024 competes against Rubin in 2027. The ASIC design cycle is 2–3 years; Nvidia's product cycle is now annual.

Net: the sourced evidence supports a workload split, not a replacement: custom ASICs win on cost for stable, owner-operated, inference-heavy workloads at hyperscale volumes; GPUs win on flexibility, time-to-market, software and rentability. That is exactly the split the customers in Section 03 are implementing.

Section 06Valuation and what would have to change

Why forward P/E disagrees so widely: the numerator is the same price; the denominator is not the same earnings. Sources differ on (a) horizon — next-twelve-months consensus vs fiscal-2027 vs fiscal-2028 EPS; (b) GAAP vs non-GAAP — decisive for Broadcom, whose GAAP EPS excludes ~$2.0B/quarter of amortization and ~$2.0B of stock compensation that non-GAAP adds back, and for Nvidia, whose GAAP EPS currently includes billions in equity-investment gains; and (c) how fast each provider rolls the explosive guided quarters (Nvidia $108B, Broadcom $34.8B) into "forward." The honest statement is: on any single consistent provider, Broadcom trades at a forward-P/E premium of roughly 1.1–1.5× to Nvidia (e.g. StockAnalysis 19.08 vs 17.08; MarketBeat ~19.0 vs ~13.2), while on trailing earnings Broadcom's premium is ~1.6× (42 vs 26) — the market is paying today for Broadcom's faster guided growth to close the earnings gap, exactly the bet Section 04 describes. This piece makes no recommendation and adopts no price target; analyst average targets exist in the sources (e.g. StockAnalysis averages of $327.14 for NVDA and $526.81 for AVGO) and are reported here only as source content, not as this piece's claims.

What would have to change — for Broadcom to materially close

ALL of the following must hold (company guidance vs analyst scenario labelled):

1. Broadcom hits its own staircase — COMPANY GUIDANCE. AI revenue $58B (FY26, incl. $21.7B guided Q4) → ~$115B (FY27) → ~$230B (FY28). Arithmetic check: $115B would be 41% of Nvidia's current Data Center TTM; $230B would be 83% of it. The staircase alone does not close the gap unless Nvidia stalls — it closes the ratio from 6.55× (TTM) toward ~2–3× only if Nvidia's Data Center growth simultaneously falls from +117% to low-double-digits, which is the opposite of Nvidia's ~70% FY28 growth guide. An actual parity scenario requires FY29+ at the same Broadcom slope.
2. The Anthropic/OpenAI leg must fund itself — ANALYST/FILING RISK. Broadcom's FY27–28 slope leans on two pre-profit labs: Anthropic (expected by Broadcom to be its largest XPU customer in 2027; ~$161B Broadcom lease in its S-1; Mizuho-estimated $42B of 2027 Broadcom revenue) and OpenAI (10 GW term sheet; $30–50B+ financing reportedly being arranged by Broadcom, unsigned as of October 2026). Broadcom's own April 2026 filing makes Anthropic's 3.5 GW consumption conditional on its "continued commercial success," and an October 2026 8-K disclosed a Broadcom backstop for custom-AI-rack leases with maximum potential liability of ~$43B (Equibles filing summary). If lab financing tightens, Broadcom's guided slope — not Nvidia's — is the one exposed.
3. Roadmap execution — COMPANY ROADMAPS. Broadcom/partners: TPU 8t (training) and 8i (inference) — Broadcom said in September 2026 it had begun production shipments of 8i; Meta's first 2nm MTIA generations (400/450/500) deploying 2027; OpenAI Jalapeño volume in 2027; Tomahawk 6 ramp (production since March 2026); Samsung 2nm/HBM supply under the July 2026 MOU coming on line (MOU stage — yields for Samsung 2nm at volume are unproven, per TechTimes coverage). Slippage in any of the 2nm ramps pushes the FY28 claim out a year.
4. ASIC share of units must convert to share of dollars. Goldman/JPMorgan unit-parity-by-2027 scenarios (Section 04) still leave Nvidia with the majority of revenue if ASIC units book at ~40% of GPU prices. Broadcom closing the revenue gap additionally requires rack-scale system revenue (it is already moving from chips to racks — the Anthropic lease structure and the OpenAI rack term sheet are exactly that) rather than silicon-only content.

What would have to change — for Nvidia to extend

The conditions are the mirror image — plus its own execution:

1. Vera Rubin ramp — COMPANY GUIDANCE. In full production as of August 2026; ~20% of Q3 FY27 Data Center revenue guided; $40B revenue per gigawatt (vs $25B Blackwell) means Nvidia can grow revenue ~60% on flat gigawatt additions. Rubin Ultra follows in 2027 and Feynman in 2028 on the stated annual cadence. The guided gross-margin trough (74% in Q3 FY27, 71–72% in Q4 FY27 on memory costs, per call reporting) is the near-term watch item: extension on Nvidia's terms requires margin recovery in FY28, not just revenue.
2. Hyperscaler capex concentration — SHARED RISK, ASYMMETRIC EXPOSURE. Combined 2026 capex for the four largest hyperscalers is estimated at ~$720–745B (~$835B including Oracle), tracking toward $1T+ in 2027 (TIKR/TradingNews syntheses of company guides). Nvidia's concentration is disclosed in its own split: Hyperscale = $48.7B of $89.0B quarterly Data Center revenue, with press analysis of its 10-Q putting the top four customers at ~61% of total revenue. Broadcom's concentration is narrower and deeper — effectively six XPU customers, with Google historically the largest. A capex pause hits Nvidia's revenue first and largest in dollars; a single-customer defection (e.g. Google dual-sourcing TPU design to Marvell/MediaTek, both reported in talks during 2026) hits Broadcom's growth rate hardest.
3. Export controls — POLICY VARIABLE, MOSTLY NVIDIA'S PROBLEM TODAY. Nvidia's outlook assumes zero China Data Center compute revenue; CFO Kress on Aug 26, 2026: H200 shipments to China-based customers were <1% of Data Center revenue in Q2, under a January 2026 BIS regime of case-by-case licences, a 25% duty and per-firm caps (~75,000 units), with Beijing separately restricting entry. Huang's own framing (company claim): Nvidia's China AI-accelerator share went from ~95% to zero. Any loosening is pure upside to Nvidia's reported base and a headwind to the China-ASIC ecosystem (ByteDance's Broadcom/MediaTek/Qualcomm programs exist because of the controls); any tightening that captures ASIC design services would newly hit Broadcom's reported ByteDance exposure — which is itself only press-reported (Section 03).
4. The software moat must hold. Nvidia extends if CUDA + NVLink Fusion keep even ASIC fleets buying Nvidia interconnect/CPU/NIC content; Broadcom closes faster if open Ethernet (Tomahawk/SUE, UALink) becomes the default scale-up fabric and TorchTPU/JAX portability erodes the CUDA switching cost. TorchTPU's reported progress (PyTorch on TPU at ~94% of JAX performance, serving costs down 44% in six weeks on software alone — SemiAnalysis via Memeburn) is the kind of evidence to watch.

Analyst scenarios, labelled as such: Counterpoint (Broadcom holds ~60% of ASIC design through 2027; ASIC shipments surpass data-center GPUs by 2028), Goldman Sachs (ASIC/GPU unit parity by 2027), JPMorgan (ASICs 53% of AI chip shipments in 2027) and Mizuho (Anthropic = $42B of Broadcom's 2027 AI revenue) are analyst models, not company guidance. No analyst scenario found in this research puts Broadcom's AI revenue at parity with Nvidia's Data Center revenue before 2029.