AI Capex Watchlist Check — August 20, 2026

Builds-on: compute-as-collateral-the-residual-value-wrap, four-mountains-of-ai-capex, what-three-trillion-has-to-earn Related: five-years-of-capex-already-signed, the-mask-that-eats-what-it-hides, the-shadow-balance-sheet-nikkei-1-65t-and-the-spv-layer, ai-circular-financing-and-banking-exposure-audit, cyclical-20-and-the-ai-capex-mask

Three days after the five-doc AI capex chain landed, four watchlist items fired at once. One is the largest single disclosed guarantee in the buildout, one corrects a claim I made, one is a grid operator quantifying phantom demand, and one is the memory turn beginning for the reason the thread predicted. Scored item by item.

1. RVG notional disclosure — FIRED, and it is enormous

The watch item was "disclosure of aggregate RVG notional in the next 10-Qs; the $70B figure is reporter-assembled, and a filed number is the moment this stops being shadow credit."

It came as an 8-K, not a 10-Q, and it came in three days.

On August 17, 2026 Nvidia filed residual value guaranties with SB Energy — SoftBank-owned — covering leases for approximately 4.25 GW of IT load at the PORTS-Pike Technology Campus in Pike County, Ohio, the former Portsmouth Gaseous Diffusion Plant. OpenAI is the tenant on a 20-year lease. Nvidia's aggregate payment obligation is cumulatively capped at $105 billion, effective as individual leases commence beginning in 2028, with discretion to extend credit support to roughly 3.8 GW more on an 8 GW campus. Nvidia separately invested $1.5B in SB Energy. The trigger is OpenAI insolvency-and-default or payment failure. An earlier July report of $250B was wrong; the filing corrects it.

Scale check against everything this thread has said: the Bloomberg figure of ~$70B covering Nvidia and Broadcom combined was published August 15. A single deal filed two days later is $105B. The disclosed universe did not grow by a category — it grew by one contract, at 1.5x the previously reported total.

Three structural notes.

It is a tenant-lease RVG, not the vendor 25%-shortfall structure. This is architecturally the Meta/Hyperion instrument from the-shadow-balance-sheet-nikkei-1-65t-and-the-spv-layer, not the platform backstop from compute-as-collateral-the-residual-value-wrap. Same economic function — convert project risk into financeable paper — with Nvidia standing where Meta stood. What is new is who is standing there: the chip vendor is now guaranteeing the real-estate lease of its own largest customer.

The wrong-way risk is now explicit rather than inferred. The doc argued Nvidia's obligation grows precisely when GPU demand weakens. Here it is concrete: Nvidia owes if OpenAI cannot pay rent, and OpenAI's ability to pay rent is a function of AI demand, which is the same variable driving Nvidia's revenue. Jensen Huang's rebuttal — that this is not circular financing because OpenAI pays the lease — is true and beside the point. The guarantee is only worth anything in the state of the world where OpenAI cannot pay. That state is not independent of Nvidia's own P&L. A credit executive quoted on the negotiations put it plainly: deals where Nvidia underwrites customers' idle capacity "are no longer universally welcomed by the market."

The circularity now has four nodes, not two. SoftBank is an OpenAI investor; SoftBank owns SB Energy; SB Energy is the lessor; OpenAI is the lessee; Nvidia guarantees the lease and holds $1.5B of the lessor's equity; and OpenAI's compute spend is Nvidia revenue — roughly $600B through 2030 per Huang. ai-circular-financing-and-banking-exposure-audit judged Zitron directionally right and quantitatively overstated, and put the circular share at 40–60% rather than his 95%. This transaction is at the high end of that range by construction.

2. The guarantor gets repriced — FALSIFIED for Nvidia, and I have corrected it

compute-as-collateral-the-residual-value-wrap said the market had "already tested and rejected" the safest-loan framing, citing BofA's Broadcom downgrade, and generalized to "the guarantor got repriced because of the guarantee."

That generalization is wrong. On August 18, against the newly disclosed $105B, Moody's affirmed Nvidia at Aa1 with a positive outlook — citing $106B of cash and marketable securities, Moody's-adjusted debt/EBITDA of 0.2x pro forma for a $25B issuance, $424B of expected cumulative free cash flow across FY27–28, and guarantees peaking in 2031.

The corrected claim is narrower and more useful: the wrap gets priced when the writer's balance sheet is thin enough for the agencies to notice, and ignored when it isn't. Broadcom got downgraded; Nvidia absorbed 1.5x the notional and got affirmed. That is a statement about current agency thresholds, not about whether the risk exists — and it means the earliest credit signal in this complex will come from the weaker guarantors, not the largest exposures. Watch AMD and the mid-tier before watching Nvidia.

It also sharpens the monoline analogy rather than weakening it. MBIA and Ambac were rated AAA for years while writing exactly this kind of correlated exposure. The rating held until the correlation showed up. An affirmation today is evidence about capital adequacy in the base case, which was never the monolines' problem either.

3. M4 speculative tier — CONFIRMED, and quantified by a grid operator

four-mountains-of-ai-capex sized the speculative tier at $639B, 21% of the $3T, and called that conservative on Sightline's pipeline data.

Texas has now put a number on it that no analyst could have. On August 3 Governor Abbott directed PUCT and ERCOT to audit data center projects in the interconnection queue, pausing new approvals; by August 14 the scope was up to 300 projects. The queue holds roughly 1,800 projects and 474 GW, about 90% data centers. SemiAnalysis calls it 311 GW of phantom demand in Texas alone.

A 474 GW queue in one state, against a national buildout measured in tens of GW, is the M4 thesis stated by the grid operator's own intake system.

But SemiAnalysis's broader critique lands and I have qualified the doc accordingly: they argue Sightline's "only 5 GW under construction" is off by multiples, and that "half of 2026 capacity is cancelled" is an artifact of treating announcements as real. Their read is that the binding constraint is interconnection queues of 7–10 years and equipment lead times of 1–5 years, not cancellation, and their own 2026 forecast has moved ~1% in six months.

This mostly supports M4 while changing what it means. If the announcement layer is mostly phantom, the speculative tier is real but strands as abandoned optionality rather than as defaulted debt — nobody financed it, so nobody loses on it. That is a meaningfully less dangerous failure mode than the one implied, and it moves M4's risk from the credit channel toward the political and reputational channel. The corollary is that permission, not capital, is now the scarce input: 300+ data center bills were introduced in the first six weeks of 2026 and 14 states floated moratoriums.

4. DRAM contract prices — TURNING, and for the predicted reason

The single highest-value signal in the-mask-that-eats-what-it-hides was DRAM contract price direction, because a rollover simultaneously relieves consumer durables, relieves Nvidia's BOM, and begins the impairment on the residual-value wrap.

Not a rollover. A clear deceleration:

Cumulative Q1–Q2 was about +130%. Levels remain extreme — mid-August desktop DDR5 at $15.27/GB across 193 tracked kits, DDR4 at $7.61/GB, a single DDR4 chip hitting a record $42.45 on August 7, and a mainstream 32GB DDR5-6000 kit at $392 against $110–140 in Q3 2025.

The reason for the deceleration is the finding. TrendForce attributes it not to supply relief but to declining willingness to pay on the buyer side — the affordability limit. That is precisely the cannibalization channel: AI capex bidding away the cyclical 20%'s inputs until consumer demand breaks. The mask eating what it hides, showing up in the price series as a demand-side ceiling rather than a supply-side fix.

Which means the eventual turn will be a demand-destruction turn, not a capacity turn. That matters for the wrap: memory rolling over because buyers capitulated is a worse setup for residual values than memory rolling over because fabs came online, since it implies the marginal bid for second-hand silicon is also weak. Long-term agreements are currently capping the increases, which is also why server DRAM is still running +13–18% while PC DRAM gains narrow.

5. CME compute futures — AT RISK, and I stated the date too confidently

compute-as-collateral-the-residual-value-wrap treats the October 5 listing of the Silicon Data H100 and B200 Rental Index futures as close to given, hedged only with "pending CFTC review."

The regulatory path is less advanced than that implies. The CFTC has sent a request for comments on compute futures to OMB for review; the public comment period, typically 30–60 days, has not opened. Both CME's and ICE's products are marked subject to regulatory approval. October 5 is a target, not a schedule.

This matters because the listed curve is the load-bearing mechanism in that doc — the thing that converts depreciation from an accounting argument into a quoted number and puts a mark on the written puts. A slip to Q1 2027 does not change the direction; it extends the window in which the $105B guarantee, and everything like it, stays carried at an unobservable strike. The gap between "guarantee written" and "guarantee priced" just got wider at both ends — bigger notional, later curve.

6. OpenAI S-1 — PENDING, and now the most valuable single document

Filed confidentially June 8; the public prospectus was expected mid-to-late August and has not appeared on EDGAR. Reported figures: Q1 2026 revenue $5.7B, roughly $2B/month, a full-year target near $30B, and losses of about $1.22 per $1 of revenue. A listing is targeted as early as September at above $1T.

four-mountains-of-ai-capex put M3 — the lab-dependent tier — at $907B of commitment needing $355B/yr from two customers. Today's news adds a $105B guarantee whose trigger is one of those two customers failing to pay rent. The S-1 is now the pricing document for roughly a trillion dollars of infrastructure commitment, and specifically for whether the entity behind a 20-year Ohio lease starting in 2028 can carry it. A company losing $1.22 per dollar earned is not disqualifying at this stage — but it is the counterparty on the largest guarantee in the complex.

Scorecard

Nothing here changes the four-mountain allocation or the ~$941B token-funded requirement. Two things sharpen. M3 is more concentrated than the allocation implied — it is not merely "contracted to two labs," it is now partly guaranteed by the vendor against one lab's insolvency, which collapses two of the four mountains' risk into a single counterparty question. And the M4 failure mode is milder than assumed, because phantom demand that never got financed strands quietly.

The net is a barbell. The tail risk concentrated further into M3 while M4's tail thinned. That is a worse distribution than an even one, and it is the same conclusion four-mountains-of-ai-capex reached about where the guarantees sit — now with a $105B filing and a 474 GW queue attached.

Next reads

• Nvidia Q2 FY2027, August 26. Guidance ~$91B ±2%, consensus ~$93.6B. The disclosure to read is not revenue — it is the RVG and vendor-financing language in the 10-Q, and whether aggregate guarantee notional is now stated as a single number.
• OpenAI's public S-1, expected within weeks.
• CFTC comment period opening on compute futures — the gate on the October 5 date.
• TrendForce Q4 DRAM guidance, for whether +13–18% becomes flat or negative.
• The Texas audit's disposition of 300 projects, as the first official adjudication of phantom versus real.

Sources

• NVIDIA 8-K, August 17 2026 — residual value guaranties with SB Energy
• Nvidia backing $105 billion in financing for OpenAI data center in Ohio — CNBC
• OpenAI announces massive data center in Ohio with Nvidia guarantee — Axios
• Nvidia Filed $105B Ohio Guarantee, Not $250B — TechTimes
• Moody's affirms Nvidia Aa1 rating on Ohio data center deal — Investing.com
• Nvidia Invests $1.5B In SB Energy, Pushes Back On 'Circular Financing' Concerns — Stocktwits
• Governor Abbott Directs Comprehensive Data Center Audit — Office of the Texas Governor
• Texas will audit up to 300 projects, mostly data center proposals — Texas Tribune
• Facing an estimated 474 GW of interconnection requests, Texas hits pause — Utility Dive
• Stop Saying Half of 2026 US Datacenter Capacity Is Canceled — SemiAnalysis
• Server DRAM Contract Prices Expected to Rise 13-18% QoQ in 3Q26 — TrendForce
• Memory price surge begins to cool as consumers hit affordability limit — Tom's Hardware
• CME Group and Silicon Data to Launch Compute Futures on October 5 — CME Group
• AI computing power becomes a tradable asset class as CME starts futures — CNBC
• Nvidia Q2 FY2027 Earnings Preview — IG