Why the custom-silicon insurgency keeps hitting the financing wall

Companion article to Issue 09 // 8 May 2026 // R. Lauritsen

Why the custom-silicon insurgency keeps hitting the financing wall

OpenAI, Broadcom, and Microsoft just exposed the structural problem with every Nvidia alternative: someone has to underwrite a programme designed to compete with their own future revenue. They won’t. Here’s what that means for AVGO, NVDA, ORCL, and CoreWeave — and the three signals that decide who wins the next twelve months.

12 min read // Sectors: AI infrastructure, custom silicon, hyperscaler capex

THE THESIS IN ONE PARAGRAPH

The custom-silicon trade hasn’t failed. Trainium ships. TPU sells externally. MTIA Gen 2 is in production. The engineering works. What broke this week is the financing model that was supposed to bring those programmes to scale fast enough to actually displace Nvidia revenue. Project Nexus exposed the rule that governs every custom-silicon programme: someone has to underwrite a chip designed to compete with their own future revenue. Microsoft just refused to do that for OpenAI. Once that rule is visible, Nvidia’s pricing power gets two more quarters of runway, four named companies sit on the wrong side of the delay, and the May 20 NVDA print becomes the binary that decides the trade.

WHAT ACTUALLY HAPPENED ON THURSDAY

The Information broke the story Thursday afternoon, citing an internal OpenAI memo. The first phase of OpenAI’s custom chip programme with Broadcom — codename Project Nexus, chip codename Jalapeno — has stalled. Broadcom told OpenAI it will only finance the 1.3 gigawatt buildout, costing roughly $18 billion, if Microsoft commits to buying 40% of the chips and renting them back to OpenAI. Microsoft hasn’t signed.

Recall the October 2025 announcement: 10 gigawatts, $500 billion in committed capex, chips before 2030. Nexus is just the first 1.3 GW out of 10. The structural problem: OpenAI doesn’t have $18 billion in cash, never mind $500 billion. The plan was for hyperscalers — primarily Microsoft, OpenAI’s largest investor — to buy the chips and rent the compute back. AVGO closed down 4.4% on the day. NVDA closed near record highs. The AVGO–NVDA spread blew out 6.7 points by the close.

Why Microsoft said no

Two reasons, both reported. First, Microsoft prefers standard data-centre designs that run any workload. OpenAI’s plan involves specialised infrastructure optimised for Jalapeno — power densities, cooling, networking topology designed around one chip. Microsoft would be locking long-lived physical assets to OpenAI’s technology bet for a decade. Second, and more fundamental: Microsoft would be funding a chip designed to reduce its largest cloud customer’s dependence on chips Microsoft also buys. The economics only work if Microsoft is convinced Nvidia’s pricing power compresses within the depreciation life of the building. Conviction Microsoft does not appear to have.

THE STRUCTURAL RULE CUSTOM SILICON KEEPS HITTING

Every custom-silicon programme needs the same three ingredients.

When all three are inside one company, the programme works. Google designs TPUs, deploys them in Google Cloud, contracts Broadcom for parts. Meta designs MTIA, deploys internally, contracts Broadcom for layout. Amazon designs Trainium, deploys in AWS, contracts Marvell. Self-contained economics.

When the model lab and the hyperscaler are different companies, the programme stalls. OpenAI is the model lab. Microsoft is the hyperscaler. Broadcom is the vendor. The vendor will only commit if the hyperscaler underwrites — and the hyperscaler’s incentive to underwrite is conditional on a future where Nvidia’s pricing power has compressed. That future has not yet arrived.

The counter-example: Anthropic uses Trainium. Amazon designed Trainium and deploys it in AWS, where Anthropic runs. Up to $25B in compute commitments. Model lab and hyperscaler aligned. No third party to convince. That programme is shipping; OpenAI’s isn’t.

WHO’S EXPOSED: THE NAMED-LOSER MAP

A two-sided thesis is stronger than a one-sided thesis. The financing wall doesn’t hurt everyone equally. Three names sit on the wrong side of the delay; three names benefit. Each placement earns its row.

Oracle: 76% capex-to-revenue is the tell

Oracle signed a $300B, five-year compute contract with OpenAI in September 2025 — anchor tenant for Stargate. The number that matters: Oracle is running a 76% capex-to-revenue ratio in FY26 ($50B capex on $66B revenue). Meta runs 30%. MSFT 38%. GOOGL 33%. AMZN 28%. Total debt $134.6B. Debt-to-equity 427%. Credit rating BBB. (Source: Oracle Q2 FY26 filings.) If the $300B contract slips 12–18 months on chip-supply or power constraints, leverage climbs above 4.5x and triggers a rating downgrade. Oracle’s short-term cash flow benefits from Nexus stalling — OpenAI keeps buying NVDA capacity from Oracle’s racks. The long-term solvency question is sharper this week than last.

CoreWeave: 67% concentrated in the customer that just said no

CoreWeave’s 10-Q discloses Customer A — almost certainly Microsoft — as 67% of full-year 2025 revenue and 70% of YTD revenue through Q3. Customer A also accounts for 40% of accounts receivable. (Source: CoreWeave Form 10-Q, Q3 2025.) OpenAI added $11.9B + $4B expansion to its CoreWeave commitment over 2025. Why this is a balance-sheet event: 96% of CoreWeave revenue is locked into multi-year contracts customers pay whether they use the compute or not. In a Nexus-restructure scenario, OpenAI may push to renegotiate compute commitments to fund the gap. CoreWeave runs on borrowed money — if its two largest customers go renegotiating, the collateral assumptions in CoreWeave’s debt covenants get tested.

AVGO short-term: time decay is the trade

AVGO is the most interesting position because the short-term setback could become a long-term win. The $18B Nexus revenue is one piece of a much larger custom-silicon book — Google TPU, Meta MTIA, parts of Trainium. AVGO is up 104% over twelve months for a reason: across multiple hyperscalers, it’s the cleanest pure-play exposure to AI capex outside Nvidia. The print question: does Microsoft sign in Q2, Q3, or push into 2027? Each quarter of delay is a multiple compression. A smaller commitment by July (25–30% rather than 40%) reloads the trade. A push into 2027 turns timing risk into structural risk.

WHY NVDA WINS ON A DELAY

The simplest version: NVDA still owns roughly 90% of accelerator share. Its Q3 FY26 disclosure showed four direct customers each over 10% of revenue — Customer A at 22%, Customer B at 15%, Customer C at 13%, Customer D at 11%. 61% of the world’s most important chip company depends on four mystery customers. But that concentration is also what makes the trade work right now: every dollar of OpenAI compute that doesn’t go to Jalapeno goes to NVDA Blackwell instead.

OpenAI is projected to spend over $200 billion in operating expenses through 2029. The chip portion of that is real money. Without a working custom-silicon programme, OpenAI defaults to buying NVDA — directly through Microsoft Azure, or indirectly through Oracle Stargate or CoreWeave. NVDA’s pricing power is durable for as long as the alternative isn’t funded.

The receivables question that decides the May 20 print

The bear case on NVDA isn’t about demand. It’s about quality of revenue. NVDA’s accounts receivable jumped to $33.391 billion in Q3 FY26, up from $23.065 billion the prior quarter. Michael Burry has called the pattern "fraud, not flywheel" — his thesis being that NVDA’s growth is partly customers paying late, partly NVDA committing to buy back its own GPUs through customer cloud capacity, and partly hyperscalers overbuilding with Nvidia balance-sheet support.

Why this matters now: the Nexus story raises exactly the question Burry has been asking — can the largest customers actually pay for what they’ve committed to? OpenAI’s $200B opex burn through 2029 is the most leveraged customer in the complex. If receivables print above $40 billion in the May 20 quarter, the Burry thesis transitions from a fringe call to a real concern. If receivables flat or down + raised guidance, the rally extends into year-end.

THREE SIGNALS THAT DECIDE THE TRADE

The whole thesis comes down to three observable events in the next ten weeks. Calendar-checkable, binary, and tradeable. None of them require interpretation — they print or they don’t.

How to use these: the May 20 print is the highest-information event because it tests both the NVDA bull case and the Burry receivables thesis simultaneously. Watch the AVGO–NVDA spread heading into it — if the spread compresses pre-print, the market has already priced in a Nexus restructure. If it stays wide, the binary is still live.

SCENARIO PROBABILITIES

These are my probability estimates, not market-implied. Calibrate them against your own view of MSFT’s incentives, OpenAI’s funding capacity, and the political calendar.

The base case (~55%) is the Bear Case Box outcome. Most likely: a smaller MSFT commitment by Q3, Broadcom resumes funding at reduced phase-one scope, the AVGO trade reloads at a slightly delayed timeline. The reason this isn’t a higher probability: it requires Microsoft to do exactly what it just declined to do, only smaller. That’s plausible but not automatic.

The right tail risk for AVGO short positions: a Microsoft signature in July would print as a major positive catalyst even with a smaller commitment. The shorter the period MSFT takes to sign, the more painful the squeeze. This is why size matters — directional conviction without sizing for the squeeze risk is what kills these trades.

WHAT THIS MEANS FOR THE FRAMEWORK

The conviction calls in the AI Investment Framework now have a single pivot point: the Nexus financing wall. Here’s how to read each layer through that lens.

Infrastructure: HIGH → HIGH ↑↑

Conviction raised. The Nexus stall is the cleanest read of the week. The alternatives to Nvidia aren’t ready to fund themselves, and that means the monopoly tax holds for at least another two quarters. NVDA, AVGO (long-term), TSM all benefit from the layer; AVGO short-term faces the financing-overhang trade discussed above.

Platforms: HIGH ↑

GOOGL is the only Western company that owns all four layers AND prints the revenue. The Nexus story makes vertical integration look better the harder triangulated programmes get to fund. MSFT and AMZN catching up but still lack the silicon. Watch for TPU disclosure in Q2 earnings as the structural re-rating catalyst.

Cybersecurity: HIGH → MEDIUM ↓

Conviction lowered. Mythos was the catalyst two weeks ago; this week, no follow-through. Back to "wait for CRWD earnings June 1." Two weeks of no incident-driven repricing means the layer reverts to its baseline thesis.

Global, Physical AI, Applications: no change

DeepSeek V4 + Huawei is the chip-and-model stack outside the U.S. — the friction is whether Western buyers can deploy Chinese silicon. Physical AI (Optimus Q2 validation event) and Applications (squeezed middle, ~$2T lost from peak) are tracking their independent narratives.

BOTTOM LINE

Last Friday’s call: capital is bifurcating away from Nvidia toward custom silicon.

This Friday’s call: still right, six months early. The custom-silicon trade hasn’t failed — Trainium ships, TPU sells externally, MTIA is in production. What broke is the financing rail that brings non-aligned programmes (model lab + separate hyperscaler + vendor) to scale. Nexus is the canary. AVGO’s multiple compresses every quarter MSFT doesn’t sign. Oracle’s 76% capex/revenue ratio works only if Stargate converts on schedule. NVDA harvests the runway.

Stay curious — and stay qualified.

— R. Lauritsen

Editor, iPrompt Signals

METHODOLOGY AND SOURCES

Reporting comes from named primary sources cited inline; estimates and inferences are tagged in italic grey throughout the piece. Primary reporting: The Information (Nexus financing snag, 7 May 2026); company filings (NVDA Q3 FY26 10-Q, CoreWeave 10-Q Q3 2025, Oracle Q2 FY26 disclosures); OpenAI/Oracle $300B contract announcement, September 2025. Estimates and inferences: scenario probabilities (my own); the three-ingredients framework synthesised across Anthropic–Amazon, Google–TPU, and Meta–MTIA precedents; the "two extra quarters" estimate for NVDA pricing-power durability.

Disclaimer: For informational purposes only. Not financial advice. iPrompt Signals is not a registered investment advisor. Conduct your own research and consult a qualified financial professional before making investment decisions.