OpenAI's $18B chip deal hit a wall. Microsoft built it

…all in a FREE Weekly newsletter.

AI & robotics investing — explained so you can actually act on it.
ISSUE 09 // Friday, 8 May 2026 // 6–9 min read

THE HOOK

Last Friday I argued capital was bifurcating away from Nvidia toward custom silicon — Broadcom, Trainium, TPU, MTIA. Six days later it hit a wall.

On Thursday The Information reported that OpenAI’s $18 billion Broadcom chip deal — the first 1.3 gigawatt phase of the programme that’s supposed to wean OpenAI off Nvidia — has stalled. Broadcom won’t fund it. Microsoft was supposed to. Microsoft said no. AVGO closed down 4.4%. NVDA closed near record highs.

What this means for your weekend research: the custom-silicon trade is delayed, not dead. NVDA’s moat just got two extra quarters. Three names sit on the wrong side of that delay — and the May 20 NVDA print decides which way it resolves. Pick one company below, spend 45 minutes this weekend, build your own tripwire.

WEEKLY SCOREBOARD

Bottom line: Nvidia rallied because the alternatives can’t pay for themselves. The "Nvidia tax forever" thesis isn’t a permanent state — it’s a financing problem disguised as a moat. Whoever solves the financing wins the next leg.

TOP HEADLINES

1. The $18 billion question that broke Project Nexus.

The Information reported Thursday that the first 1.3 GW phase of OpenAI’s Broadcom chip programme — codename Jalapeno, $18B in committed capex — has stalled. Broadcom won’t fund the buildout unless Microsoft commits to buying 40% of the chips and renting them back to OpenAI. Microsoft hasn’t signed. The deal isn’t dead, it’s blocked. AVGO −4% intraday on the report. What it means: the custom-silicon thesis only works if hyperscalers underwrite their own competition. They’re not — at least not on these terms.

2. Investors are repricing AI capex as cost, not growth.

Last week’s Big Tech earnings reactions are still rippling. GOOGL +10% on $460B cloud backlog and clear AI-to-revenue translation. META −8 to −9% on raised capex ($125–145B) and no cloud business to monetise it. MSFT −4% on cost concerns despite Cloud crossing $50B run-rate. AMZN essentially flat on AWS at +28% YoY. What it means: the selectivity is sharper than at any point this cycle. Investors will reward AI revenue. They will not reward AI spending.

3. AVGO–NVDA spread blew out 6.7 points on Thursday alone.

The cleanest market read of the week. Same story, opposite reactions: AVGO down 4.4% as the Nexus financing snag broke, NVDA up 2.3% as the alternative to NVDA stalled. The pair-trade desks ran with it — block prints showed concentrated AVGO selling against NVDA buying through the afternoon session. What it means: the market priced this as a relative trade, not an absolute one. The custom-silicon thesis didn’t die — its timeline got repriced, and the spread is the cleanest expression of that.

4. DeepSeek V4 ships open-source on Huawei silicon.

A genuinely under-covered story. China’s DeepSeek released V4 on April 24 — open-source weights, deeply integrated with Huawei’s Ascend chip stack. Reasoning and agentic upgrades, lower token costs. The strategic point: there is now a chip-and-model stack outside the U.S. that requires zero Nvidia GPUs, zero Western cloud infrastructure, and zero Western capital. What it means: the bear case for the Nvidia premium just got a fresh structural data point. Whether Western buyers can deploy Chinese chips is a separate question — for now, no.

5. Nvidia + Corning $3.2B optical fibre deal.

Less attention than it deserved. Nvidia announced an investment of up to $3.2 billion in Corning to lock in optical fibre supply for AI data centres. Companion story: Super Micro guided Q1 revenue more than doubling. What it means: the next bottleneck isn’t compute. It’s fibre, racks, cooling, and power. The physical AI layer is where capacity is being booked twelve to eighteen months ahead — and it’s getting overlooked because it doesn’t fit the GPU narrative.

OUR INVESTING ANGLE

Everyone’s watching whether NVDA can hold $200. The smarter watch is whether the alternatives can fund themselves.

Last Friday’s thesis assumed the buyers of those alternatives could pay for them. This week, they couldn’t. Here’s how my view changes:

The thesis: the Nvidia "monopoly tax" is more durable than last week’s bull case for custom silicon implied. Not because the engineering is wrong — because the financing is. Hyperscaler co-investment was supposed to underwrite each custom programme; this week proved that’s now conditional, not assumed. Margins hold for at least another two quarters.

So yeah. NVDA closed up 2.3% on the week into news that should have hurt it. The market read that.

Who gets hurt? Three names, not vibes. Oracle — sitting on hundreds of billions of OpenAI compute commitments funded by debt. The debt-to-AI-revenue ratio is the cleanest tell of who’s overextended. CoreWeave — the GPU-as-a-service play whose entire business depends on the OpenAI/MSFT/AVGO triangle holding. The Nexus snag is a balance-sheet event for them. AVGO short-term — the $18B revenue overhang doesn’t go to zero, but the print rides on whether MSFT signs in Q2, Q3, or 2027. Each quarter of delay is a multiple compression.

→ Read the deep dive: Why the custom-silicon insurgency keeps hitting the financing wall

THREE IDEAS TO RESEARCH THIS WEEKEND

Not recommendations — starting points for your own research. One primary, one alternative, one watchlist add.

Idea 1 — Nvidia (NVDA): the primary research target this weekend.

I’ve written the bull case on NVDA twice in six weeks. It stays in this week because Thursday should have hurt it. It didn’t. That’s the tell. If you only research one name from this issue, make it this one — May 20 earnings is the binary event for the whole thesis.

Why now: Project Nexus stalled on financing, not engineering. NVDA closed the week up 2.3% with the alternatives moving further away, not closer.

The case: Forward P/E 23.8 on FY27 estimates — cheap by its own ten-year average of 61.7. The "Nvidia is overextended" narrative needed a working alternative; this week, the alternative needed Microsoft and didn’t get it. (Source: company filings + my read of the multiple.)

The risk: Customer receivables at $33B is the warning Burry’s been calling out (reported by him publicly in March). The Nexus story raises exactly that question about OpenAI’s funding capacity.

Tripwire: Receivables above $40B in the May 20 print + any explicit mention of customer payment delays = Burry thesis becomes real. Receivables flat or down + raised guidance = rally extends into year-end.

How to research: Ticker NVDA. SMH (~16% NVDA weight) for diversified semis exposure. Most recent 10-Q is the place to start.

Idea 2 — Alphabet (GOOGL): the alternative if you want a compounder, not a binary.

I keep coming back to this one. GOOGL is the rare case where you can buy AI infrastructure (TPU), platform (Vertex), application (Gemini), and distribution (Search, YouTube, Android, Chrome) in one ticker — earning 36% margins. After Q1, the cloud backlog is $460B (company disclosure, Q1 2026 earnings call). Wall Street still prices this like an ad business with an AI option. It’s the other way round.

Why now: GOOGL +10% on April 28 earnings is barely worked through the multiple. TPU sales to external customers — confirmed last week — is the structural shift the Street hasn’t fully repriced.

The case: MSFT and AMZN lack the silicon. META has the silicon but no cloud. NVDA has the silicon but no application layer. Only GOOGL has all four.

The risk: ad cyclicality. If a 2026 oil-price-driven recession lands, the ad business compresses faster than cloud expands.

Tripwire: TPU revenue breaking out as a separate line item in Q2 earnings (mid-July) = "compete with Nvidia" becomes investable. No disclosure = it’s still just a backlog story.

How to research: Ticker GOOGL. Compare against MSFT and AMZN on relative cloud positioning, not on aggregate AI capex.

Idea 3 — Japanese industrial robotics: a watchlist add, not a Monday trade.

Honestly? I’m not entirely sold on the timing. The thesis is structural: while the U.S. argues over Optimus serial numbers, FANUC, Keyence, and Yaskawa are shipping into BMW Spartanburg right now — at half the multiple. Add to your watchlist; the entry isn’t this week.

Why now: Tesla Optimus production starts Q2 at Fremont. Validation event in real time, but priced into U.S. names already.

The case: FANUC (6954.T), Keyence (6861.T), Yaskawa (6506.T) trade at ~28x forward (Bloomberg consensus). U.S. robotics names trade at 58x. Same secular trend, half the multiple, twice the operating history.

The risk: Tesla execution. If Optimus actually ships in Q2/Q3, the humanoid trade resets and the Japanese names re-rate down with it.

Tripwire: First Optimus units leaving Fremont with serial numbers (not demos) by end of August = trade is real. Or BYD/Huawei announces an industrial robotics product = the Japanese discount becomes structural, not a mispricing.

How to research: BOTZ (43.8% industrials, heavy Japan exposure). FANUC (6954.T), Keyence (6861.T) directly.

AI INVESTMENT FRAMEWORK

Living portfolio framework by layer. Not financial advice — research starting points only.

Per-layer notes

Infrastructure — still the only layer where capacity is sold out twelve months ahead. NVDA at $211.50 (FY27 PE 23.8). AVGO at $424 — Nexus snag re-introduces financing risk short-term. TSM is the foundry monopoly under the foundry monopoly.

Platforms — GOOGL is the only Western company that owns all four layers AND prints the revenue. Cloud backlog $460B. MSFT and AMZN catching up but lack the vertical integration.

Applications — the squeezed middle. Software has lost ~$2T from peak as AI disruption arrives faster than monetisation. PLTR at 94x. CRM uncertain. NOW compressed. No conviction change.

Physical AI — Optimus Q2 is the validation event the whole layer trades on. Figure 03 at BMW Spartanburg. BOTZ at $39.88 — 43% industrials weight is fine until Optimus actually ships.

Cybersecurity — no follow-through this week kills the Mythos premium. Back to "wait for CRWD earnings June 1." Two weeks of no incident = layer reverts to its baseline.

Global — DeepSeek V4 plus Huawei is the chip-and-model stack outside the U.S. Western buyers can’t deploy it yet. That’s the friction the discount lives on.

What we’re watching (next 2 weeks)

Changes this week

● Conviction on Infrastructure raised from HIGH to HIGH ↑↑ — 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.

● Conviction on Cybersecurity lowered from HIGH to MEDIUM ↓ — Mythos was the catalyst two weeks ago; this week, no follow-through. Back to "wait for CRWD earnings June 1."

● No tickers added or removed.

Disclaimer: This newsletter is for informational and educational purposes only and does not constitute financial advice. iPrompt Signals is not a registered investment advisor. Always conduct your own research and consult a qualified financial professional before making investment decisions.

YOUR MOVE

The week’s three takeaways, traceable to the headlines and ideas above:

1. The custom-silicon thesis isn’t dead — it’s harder to fund than anyone admitted. Project Nexus shows the dependency: hyperscalers have to underwrite the rollouts that compete with them. They won’t. Not yet.

2. The market priced this as a relative trade, not an absolute one. AVGO–NVDA spread blew out 6.7 points Thursday. The bull case for AVGO didn’t die — it got delayed.

3. NVDA earnings on May 20 is the next test. Watch receivables and forward guidance. That print either confirms or breaks the monopoly tax thesis for Q2.

Now research one. NVDA is the primary target — May 20 earnings is the binary event that resolves this issue’s whole thesis. GOOGL is the alternative if you want a compounder over a binary. Japan is the watchlist add. 45 minutes, most recent 10-Q, build your own tripwire.

Hit reply and tell me which one you’re digging into. The replies are the part of this job that actually keeps me curious.

→ Read the deep dive: Why the custom-silicon insurgency keeps hitting the financing wall

Stay curious — and stay qualified.

— R. Lauritsen

Editor, iPrompt Signals

Know someone building an AI position? Forward this — they’ll thank you by Friday.

P.S. — Last Friday I sold you the bull case for custom silicon. Six days later it hit a financing wall. Directionally still right. Six months early. That’s the lesson I keep relearning at the edge of every thesis — the "what" is the easy part. The "when" is where almost everyone, including me, gets it wrong.

QUICK GLOSSARY

ASIC — Application-Specific Integrated Circuit. Think of an Nvidia GPU as a Swiss Army knife — does many things, well. An ASIC is a single razor-sharp blade — does one thing, brilliantly, and can’t do anything else. Trainium, TPU, MTIA, and OpenAI’s Jalapeno are all ASICs.

Capex — Capital expenditure. Money spent on long-lived assets like data centres, servers, and chips. Doesn’t hit profit immediately — gets depreciated over years. Spending $145B today shows up as $25–30B/year of cost over 2027–2031, not all at once.

Custom silicon — Chips designed by the buyer (Amazon, Google, Meta, OpenAI) instead of bought from Nvidia. Cheaper at scale once you’ve already amortised a billion dollars of design cost — which is why only the hyperscalers can play.

Hyperscaler — The four cloud giants who buy most of the world’s AI chips: Microsoft Azure, Amazon AWS, Google Cloud, Meta (for in-house use). Combined 2026 capex now ~$700B — roughly the GDP of Argentina.

Monopoly tax — Informal term for the price premium Nvidia charges because there’s no real alternative for AI training and inference. Roughly 70% gross margin on H100/B100/B200 GPUs. The custom-silicon programmes are designed to escape it.

Project Nexus — Internal name for the OpenAI–Broadcom custom chip programme. First phase ($18B, 1.3GW) stalled this week because Microsoft refuses to commit to buying 40% and renting back to OpenAI.

Receivables — Money owed by customers but not yet paid. Rising receivables = customers slow to pay = revenue quality concern. Burry’s been pointing at NVDA’s $33B figure since March.

Tripwire — A specific, observable event that would change a thesis. Calendar-checkable. Not "watch for news."

VIX — Volatility index. Below 17 means traders are calm. Above 25 means real fear is being priced. At 17.08 right now — calmer than the Thursday Nexus headline would suggest.