Why it matters
OpenAI has a route around Nvidia with chips it designs, systems Broadcom supplies and standard Ethernet. The 10 rests on a term sheet with no price attached. Eight months in, the only contractual number is 1.3 GW, and none of the Broadcom filings the site checked names OpenAI.
Tap a figure for the line in its source.
OpenAI gives Broadcom
- payment, never disclosed; every dollar figure in circulation is an outside estimate
Broadcom gives OpenAI
- of OpenAI-designed chips in Ethernet-wired racks, under a term sheet, by end of 2029
- the only part called a 'contractual commitment', to deploy in 2027
What happened
OpenAI designs its own AI chips. Broadcom turns those designs into working chips, supplies the networking and deploys the racks. The racks use standard Ethernet in place of a proprietary interconnect. The announced plan was for 10 gigawatts of capacity to arrive between late 2026 and 2029. Broadcom's CEO later said the capacity would arrive in 2027–2029, with no deliveries in 2026. Neither company has disclosed the cost. Every dollar figure in circulation is an outside estimate, and the estimates measure different things. Eight months after the announcement, Broadcom described a contractual commitment for 1.3 of the 10 gigawatts. None of the Broadcom securities filings the site checked mentions OpenAI. That includes its annual report, the three quarterly reports through the one filed in June 2026, three earnings releases and its 2026 proxy. The first chip, Jalapeño, was unveiled in June 2026.
Sources 7 documents — 4 company releases · 3 transcripts
Research trail Found 106 documents and read 98 of them while checking this deal and related ones.
- States 4 figures: 10 GW deal scope · instrument · fabric · announced schedule
- States 2 figures: first chip tapeout · initial deployment target
- States 2 figures: 1.3 GW contractual tranche · silicon status
- Cited by the record; no single figure is quoted from it.
- Cited by the record; no single figure is quoted from it.
- Cited by the record; no single figure is quoted from it.
- Cited by the record; no single figure is quoted from it.
The record's own summary of its sourcing: OpenAI/Broadcom joint release 10/13/2025 + Jalapeño release 6/24/2026 + AVGO Q4 FY25, Q1 FY26, Q2 FY26 calls · as of Jun 2026
The evidence 8 sourced figures · 6 disputes
The record's headline OpenAI–Broadcom: 10 OpenAI-designed accelerators on Ethernet (contractual so far: 1.3 GW in 2027)
- Size
- 10 GW; $ undisclosed
- When
- Oct 2025 → racks targeted 2H 2026–2029
- Status
- announced; first silicon delivered (Jalapeño, Jun 2026)
- Structure
- custom-ASIC co-development + supply agreements, rack-deployment term sheet, Ethernet scale-up/scale-out fabric
- Programme
- OpenAI–Broadcom
- Parties
- OpenAI accelerator + system designer, deployer · Broadcom silicon implementation + Ethernet/PCIe/optical fabric + rack deployment
- Record
- disclosed / primary source · as of Jun 2026
Flows 1 as the record states them
- OpenAI → Broadcomchip-spend10 GW of custom-accelerator + networking racks; $ never disclosed; only 1.3 GW (2027) called a 'contractual commitment'
Resolved figures 8 each with the verbatim line from its document
- deal-scope10 GW10 GW of OpenAI-designed custom accelerators, developed and deployed with Broadcom — the headline scope, stated by both parties in one joint release and restated by Broadcom's CEO on the Dec-2025 and Jun-2026 calls. No Broadcom 10-K/10-Q/8-K on file names OpenAI or the 10 GW (see contested binding-status)
“OpenAI and Broadcom Inc. (NASDAQ: AVGO) today announced a collaboration for 10 gigawatts of custom AI accelerators. OpenAI will design the accelerators and systems, which will be developed and deployed in partnership with Broadcom.”
OpenAI/Broadcom joint release (Broadcom IR copy) · investors.broadcom.com - instrumenttwo pieces of paper: pre-existing 'long-standing agreements' on accelerator co-development and supply, plus a signed TERM SHEET for the rack deployment — the 10 GW is the term-sheet half. The release names the instruments and never characterizes enforceability
“OpenAI and Broadcom have long-standing agreements on the co-development and supply of the AI accelerators. The two companies have signed a term sheet to deploy racks incorporating the AI accelerators and Broadcom networking solutions.”
OpenAI/Broadcom joint release · investors.broadcom.com - fabricthe racks are 'scaled entirely with Ethernet' — Broadcom's standards-based scale-up and scale-out networking in place of a proprietary interconnect — and go into OpenAI's own facilities and partner data centers. The 'routing around Nvidia' reading rests on this sentence; the release itself never names Nvidia
“The racks, scaled entirely with Ethernet and other connectivity solutions from Broadcom, will meet surging global demand for AI, with deployments across OpenAI’s facilities and partner data centers.”
OpenAI/Broadcom joint release · investors.broadcom.com - announced-schedule2029racks 'targeted to start in the second half of 2026, to complete by end of 2029' — the window as released in Oct 2025. Broadcom's CEO re-cut the start eight weeks later ('Not 26. It's more like 27, 28, 29'), so this is the announced schedule, not a confirmed one; see contested deployment-window
“Broadcom to deploy racks of AI accelerator and network systems targeted to start in the second half of 2026, to complete by end of 2029”
OpenAI/Broadcom joint release · investors.broadcom.com - contractual-tranche1.3 GW1.3 GW to deploy in 2027 is the only slice either party has called a 'contractual commitment' — Broadcom's CEO, prepared remarks, 'as part of the larger 10-gigawatt … agreement'. Supersedes his own 'over 1 gigawatt' of Mar 2026 (compatible, not conflicting; TrendForce's '1 GW' is a Reuters summary of that same call). Says nothing about the other 8.7 GW
“We have a contractual commitment to deploy 1.3 gigawatts in 2027 as part of the larger 10-gigawatt that 2029 agreement we announced last year.”
Broadcom Q2 FY26 earnings call, Jun 3 2026 (Motley Fool transcript; Globe and Mail rendering identical) · fool.com - silicon-statusBroadcom says it has 'delivered silicon' and is 'on track for production late 26' (Jun 3 2026); three weeks later both parties unveiled Jalapeño with engineering samples 'running ML workloads in the lab at production target frequency and power'. Achieved: silicon in hand. Not achieved as of Aug 2026, as far as the documents on file show: production, deployment
“For OpenAI, we have delivered silicon and we are on track for production late 26.”
Broadcom Q2 FY26 earnings call (Motley Fool transcript) · fool.com - first-chip-tapeout9Jalapeño — OpenAI's first 'Intelligence Processor', an LLM-inference accelerator — went from initial design to manufacturing tape-out in nine months, by OpenAI's own account the fastest ASIC cycle in high-performance semiconductors (the Broadcom IR copy softens 'we believe' to 'what may be'). The '~18 months' in circulation is how long the two had been working together at the Oct-2025 announcement — a different clock. Celestica is named as board/rack/system partner; no primary states a die size or process node
“Jalapeño was co-developed from initial design to manufacturing tape-out in just nine months, and the custom AI accelerator program represents what we believe to be the fastest ASIC development cycle ever achieved in high-performance advanced semiconductors.”
OpenAI Jalapeño release, Jun 24 2026 (Wayback snapshot) · web.archive.org - initial-deployment-target2026'designed for initial deployment by the end of 2026' — both parties' stated target as of Jun 2026, a target and not an achieved date. Consistent with Kawwas's Oct-2025 'in production in less than 12 months from today' and Tan's 'production late 26'; volume is a separate, later milestone (1.3 GW in 2027). Tan's quote in the same release names Microsoft as a deployment partner 'beginning in 2026'. Third-party slip reports sit in contested deployment-window
“Jalapeño is the first step in a multi-generation compute platform designed for initial deployment by the end of 2026 and expanding in the years ahead, combining OpenAI-designed accelerators with Broadcom silicon implementation, networking, and connectivity technologies; and Celestica’s board, rack, and system expertise.”
OpenAI Jalapeño release, Jun 24 2026 (Wayback snapshot) · web.archive.org
Contested figures 12 not reconciled — every position side by side, with its source and tier
Scope note what this record leaves to other records
Excluded as not-this-deal: the Sep-2025 $10B XPU order (Anthropic's — Kawwas on CNBC: 'He has not given me that P.O. yet'; Tan named Anthropic Dec 2025; see goog-anth-tpu); the 3 GW figures (Meta MTIA through 2028 → meta-mtia-broadcom; Anthropic 2027 demand); the $35B Apollo/Blackstone tranche and >20 GW XPU platform (its first transaction facilitates Anthropic's expansion) and the Q2 FY26 10-Q's $29B lease backstop (counterparties unnamed in the filing) → anth-fluidstack; Broadcom's RPO/backlog, AI-revenue, customer-concentration and CoWoS-share series (entity-level); the ~26 GW OpenAI aggregate (analyst arithmetic carrying Dell'Oro's own overlap warning) and OpenAI's revenue/valuation (entity)