The Reached and the Unreached

On August 17 the Journal put out a number. The nine largest tech companies carry roughly $3 trillion of AI commitments that sit outside their financial statements ○ — $1.2 trillion in leases that have not commenced plus $1.9 trillion in purchase obligations, about five times the $600 billion of capex they reported over the trailing year. The uncommenced-lease portion is roughly four times what was disclosed a year earlier, and the total is about three times what these companies carry in existing leases and long-term borrowings.

In the same earnings season a story pointing the other way arrived. It is the argument the tech investor Kevin Baker laid out (I came across it via the Korean channel 안될공학), and it runs like this — the hyperscalers are today collecting on GPU contracts locked cheaply in 2024 and 2025, and only when those contracts expire and reprice at current rates does the real margin show up on the books.

The market seems to read these two as pessimism versus optimism. But laying the material side by side, they do not look like opposing claims so much as the front and back of one sentence. The $3 trillion of commitments exists because of the repricing expectation, and the repricing has to land for that $3 trillion to be justified. The bull is reading this bet's income statement; the bear is reading the same bet's balance sheet.

And between them sits one decisive lag. The $3 trillion and its funding cost are a proposition that has already arrived — it ran on the Journal's front page, it has been quantified in rates (BofA, +0.3pp), and a downgrade has landed. The repricing, by contrast, is a proposition that has not arrived yet — the renewal window itself has not opened. This piece is written along that lag.

The bull ledger — renewal as the trigger

Start with the mechanism. Unlike general-purpose cloud, GPU capacity for AI training is booked years ahead at the cluster level, so the 2024–25 contracts are locked at prices that do not reflect today's supply and demand. The equipment is already installed and depreciating, so if the selling price alone rises on the same equipment, most of the increment drops to operating income. A price increase on top of fixed cost — structurally, a simple story.

This quarter's readings do support the argument's premise, that demand is running ahead of supply. The big three cloud businesses accelerated together — Google Cloud +82% ◎, Azure +43% ◎, AWS +37% ◎ — and the AWS contracted backlog reached $496 billion ◎, up $132 billion in a single quarter. Amazon's management said capacity is monetizing as fast as they install it and that they could grow faster still if supply allowed (2Q26 call ◎). The remark that a substantial share of 2028 capacity is already spoken for comes from the same call.

That much is a story with data behind it. But the argument only holds if supply is still short when the renewals come — and I will come back to that premise.

The bear ledger — the cost of the same bet has already landed

What was stacked up to capture that future revenue is the $3 trillion of commitments above. And the cost side of this bet has already started printing.

The individual figures are heavy. Alphabet's purchase and contractual obligations stood at $811 billion at the end of June ○ — up $479 billion in one quarter from $332 billion three months earlier. Meta's disclosed uncommenced leases alone come to $347 billion ○, about four times the debt on its balance sheet ($83.7 billion), and Oracle is at $273.3 billion ◎, of which the lease portion ($260 billion) has grown twenty-three-fold since November 2023. In July, S&P cut Oracle to one notch above junk (BBB-) — I am recording it as this cycle's first ratings casualty.

The funding cost is sharper in the bond market. Investment-grade issuance is running near $1.5 trillion this year (+36%) ○, on pace to pass the 2020 record, and by Nomura's estimate big tech's share alone is equivalent to about 25% of the Treasury's net issuance of notes and bonds to private investors. BofA estimated that this surge in corporate and mortgage-backed issuance pushed 10-year yields up by about 0.3 percentage point this year (August note ○). BlackRock selling $12.5 billion of data center bonds at 7.53% sits on the same line.

And the thing that catches me most is this. Contrary to the usual account that these liabilities were "transferred" to private credit, the contract terms show the risk coming back. Meta has provided residual value guarantees on leases with a threshold of about $28 billion ◎, and in the Anthropic TPU deal Broadcom makes noteholders whole if the lease payments fail. The money went out, but the residual value risk has returned to the issuer in the form of a guarantee — which means the market will not take this asset without a guarantee, and I read that as a heavier signal than the bubble-warning pieces.

What connects the two ledgers — HBM, again

Back to the bull ledger's premise. For renewals to go up, supply has to still be short at that point. If supply loosens, renewals go down instead.

And when you trace why supply does not loosen, you arrive at the place this series keeps arriving at. HBM sets the ceiling — there was reporting that Nvidia went as far as weighing a cut to Rubin Ultra's memory load, which amounts to a vendor conceding the bottleneck in a product spec — and fab capacity ordered today will not produce meaningful output before 2028 at the earliest. So the hidden premise of the repricing case is precisely HBM scarcity. The two ledgers are not telling different stories; they are standing on the same bottleneck.

The weak link — who pays the higher price

This is the part that sits uneasily with me. For repricing to land, the counterparty at renewal has to be able to pay the higher price — and roughly half of the $2 trillion of long-term contracted backlog is concentrated in two companies, OpenAI and Anthropic, both deeply cash-flow negative today.

There are only two ways they carry a more expensive renewal. Convert to self-sustaining revenue (unproven so far), or raise more. And the funding environment is tightening exactly as described above — rates are up, spreads are wider, and the ratings casualties have started. The tighter funding gets, the shakier the party that pays for the repricing becomes: the two ledgers bite each other.

Operating leverage cuts both ways here. Depreciation is fixed, so profits explode if prices rise — and for the same reason losses explode if prices fall. Fixed cost amplifies a direction; it does not choose one.

The current drawdown is the price of what already arrived

Someone will say the market has already voted, so let me write this down. The recent decline looks like a win for the bear ledger, but pulled apart, every input to it comes from the side that has already arrived. Rates (10-year in the 4.7s, the 30-year at its highest since 2007), spreads, the downgrade — these are receipts for funding cost, and all three are public information already being priced. The verdict on the revenue side, the renewal rate, has not arrived.

Two sides with a lag between them are not symmetric in a market: put a confirmed cost next to a future conditional gain and, in a rising-rate stretch, the latter gets discounted harder. And this week's calendar packs the FOMC minutes, a 20-year auction, Nvidia's earnings (Aug 26) and Warsh's first Jackson Hole keynote (Aug 28) into ten days, which is the kind of stretch where positions come down first and the logic gets attached afterward.

There is an asymmetry visible here. Betting on an extension of the current decline is betting on an extension of what everyone already knows — you are shorting after seeing the Journal's front page, the BofA note and the S&P downgrade. Repricing, by contrast, is the side where something not in the price enters it the moment it lands. There is no guarantee that everything which arrived is fully reflected (if renewals go down, the $3 trillion becomes fresh downside). But if the question is which seat is better — chasing the tail of known bad news, or waiting on a verdict nobody has seen — I am on the side that prepares for the latter.

Preparation is observation, not hope

So the conclusion of this piece is: do not bet on the decline; prepare for the side that has not arrived to land. But for preparation not to collapse into hope it needs content, so here is the content.

The content is watching the verdict indicator. The whole loop above converges on a single observable point.

At the renewal window opening from the second half of 2026, at what rate do the 2024–25 vintage GPU clusters renew?

If it goes up, repricing, AI revenue catching up with capex, and GPU residual value are all proven at once. A previous-generation cluster selling for more means that GPU's economic life got longer, which also settles the depreciation argument (a real life of 2–3 years versus a 6–7 year waterfall). What enters the price then is what is not in the price now.

And here is the condition under which this preparation is wrong. If the renewal rate goes down — meaning supply loosened, or the paying party (OpenAI, Anthropic) buckled — then the $3 trillion meets falling revenue and the fixed-cost leverage runs in reverse. If that signal comes first, the right move is to fold this preparation. As a leading proxy I will watch GPU rental spot prices (tracking), and as secondary indicators the next funding terms for OpenAI and Anthropic and AI bond spreads.

Since the standard is being written down before the renewal window opens, this piece goes on the stand when the result comes. A record that stands whether it was right or wrong is part of how this series is designed.

A note on differentiation. If repricing holds, the largest beneficiaries are in order of installed base (Microsoft, Google, AWS) — and if it fails, so is the largest damage; fixed-cost leverage runs both directions along the same ranking. The three memory makers (Samsung, SK hynix, Micron) hold the scarcity that is the source of renewal bargaining power either way, and Oracle is already paying for it in its rating. The same renewal window will be booked to a different account at every name.

Sources

Off-balance-sheet commitments (WSJ analysis, 2026-08-17)

• ~$3T across nine companies = $1.2T uncommenced leases + $1.9T purchase obligations · WSJ 2026-08-17, "Why Big Tech's AI Spending Is $3 Trillion Higher Than It Seems", analysis of filing footnotes ○ · licensed reprint · the nine = Alphabet, Amazon, Microsoft, Oracle, Nvidia, Broadcom, SpaceX, AMD, Meta
• Trailing twelve-month capex ~$600B · same analysis ○
• Uncommenced leases ≈ 4x a year earlier / total ≈ 3x existing leases plus long-term borrowings · same analysis ○
• Alphabet purchase and contractual obligations $811B (June 30), $332B three months earlier · same analysis ○
• Meta uncommenced leases $347B (as of June), on-balance-sheet debt $83.7B · same analysis ○ · a separate, broader tally that adds the CoreWeave take-or-pay and purchase obligations ($420B) uses a different definition and is not used in the body

Per-company obligations and ratings

• Oracle $273.3B = $260B uncommenced leases + $13.3B unconditional purchase obligations · Oracle FY2026 10-K (as of 2026-05-31) ◎
• Oracle lease portion up ~23x (+2,239.62%) since November 2023 · filings-based tallies (Hudson Labs, Calcbench) ○
• S&P cuts Oracle to BBB- (2026-07-09) · S&P Global Ratings ◎ · multiple reports ○

Cloud readings (2Q26)

• Google Cloud +82% ($24.8B), cloud operating margin 35.6% · Alphabet 2Q26 results ◎
• Azure +43% (above 39–40% guidance) · Microsoft FY26 Q4 8-K ◎
• AWS +37% ($42.2B), fastest in 18 quarters · Amazon 2Q26 8-K ◎ · CNBC ○
• AWS backlog $496B, up $132B from $364B the prior quarter · Amazon 2Q26 results and call ◎ · tally attributing $100B of the increase to the Anthropic contract ○
• "Monetizing as fast as we install / could grow faster with more supply" · Amazon 2Q26 earnings call ◎

Funding cost

• Investment-grade issuance ~$1.5T (+36% YoY), on pace to pass the 2020 record · Bloomberg, 2026-08 ○
• Nomura: big tech borrowing ~$200B ≈ 25% of Treasury net issuance of notes and bonds to private investors, 5x its 2025 share · Nomura estimate ○
• BofA: the surge in corporate and MBS issuance raised 10-year yields by ~0.3pp this year · BofA economists' note ○
• BlackRock sells $12.5B of bonds at 7.53% (2026-07-27) · Bloomberg ○ · El Paso, Texas campus project; GIP and HPS hold 80%, Meta 20%

Risk that came back

• Meta residual value guarantee threshold ~$28B in aggregate (leases commencing 2029, no liability recorded) · Meta quarterly report as of 2026-06-30 ◎ · distinct from the same venture's ~$12.3B initial lease commitment and $46B maximum exposure to loss
• Broadcom covers 100% of any shortfall to A1 and A2 senior noteholders in the Anthropic TPU deal · Broadcom securities filing ◎ · Bloomberg 2026-06-02 ○

Bottleneck and counterparties

• Nvidia reported to be weighing a lower memory load for Rubin Ultra (288GB → 192GB option) · multiple reports, 2026-08 ○
• Fab capacity ordered today yields meaningful output in 2028 at the earliest · industry capacity lead-time tallies ○ · not confirmed as any single company's management statement, so the attribution is left broad
• About half of the ~$2T contracted backlog at the four largest clouds sits with OpenAI and Anthropic · multiple tallies ○

Calendar and rates (at publication)

• 10-year 4.706%, 30-year 5.285% (2026-08-18 close). The 30-year traded above 5.33% intraday, its highest since 2007 · CNBC, US Treasury daily rates ◎
• Nvidia 2Q FY27 results 2026-08-26 (after the close) · company calendar ◎
• Jackson Hole symposium 2026-08-27 to 29; Warsh's first keynote the morning of Aug 28 · Kansas City Fed ◎

Left out of the body

• Three items — Google "using more than 50% above commitment", Microsoft "six weeks early to power-on", and "40% efficiency" — could not be confirmed against primary remarks and were not included.
• A June estimate of $1.8 trillion in off-balance-sheet commitments comes from a different compiler, company set and component definition than the WSJ tally, so no growth figure was constructed by subtracting the two.

Related

• On the funding structure: "The End of Building With Your Own Cash" (Silicon Nexus, 2026-07-19)
• On hyperscaler capex: "The Answer of the Other Four" (Silicon Nexus, 2026-07-23)

※ ◎ = primary source such as a filing or earnings call transcript; ○ = reporting or estimate. Accounting items differ by definition depending on the tally (uncommenced leases / purchase obligations / RPO), so the original context is given alongside. Rates and calendar items are as of publication and will move.

This piece is for information only and is not investment advice. Figures were re-confirmed against primary sources before publication, and estimate- or report-based figures are marked as such in the text.