The Value of Compute

First, a correction. #14 wrote "in Chicago." That was only half right. The market's creators are indeed Chicago-based CME Group and Silicon Data, an index provider backed by Chicago trading firm DRW. But the listing exchange is NYMEX in New York, and the product is not yet listed. Confirmed facts:

If futures are unfamiliar, think of a contract traded now for GPU rent in a future month. It is like a farmer selling the wheat price before harvest, with one month's H100 rent in place of wheat. Compute differs from wheat in one crucial way: it cannot be stored. Unused GPU time disappears rather than accumulating in a warehouse. Futures on an unstorable commodity resemble electricity futures more than oil futures. That much fits #14's intuition.

Our observations added a qualification. Fuel costs form a floor under electricity futures. For compute, calculating the electricity cost of one GPU hour (rated power × data-center efficiency × industrial electricity tariff ÷ utilization, with all assumptions stated) gives H100 USD 0.04–0.21 per hour, versus a rental quote of USD 2.79: roughly 3.3% of rent. For B200, it is 2.2%. Electricity is not the floor under compute prices. Prices are set by capital costs and depreciation: how quickly chips costing millions of won become outdated within a few years. These futures are cousins of electricity futures, but their prices will follow GPU supply and generational replacement rather than electricity. Rapid replacement also ages the benchmark: the H100 index will track an old generation within a few years. Electricity has no equivalent problem.

The exchange's official explanation is that it provides a hedging tool, which is true. But commoditization does more than hedge. In financial terms, it does three things.

⑴ Creates an asset class. Until now, two channels funded AI infrastructure: equity (hyperscalers) and credit (first-half corporate bonds of USD 225 billion, +974% year on year; USD 1.65 trillion in off-balance-sheet debt such as leases and GPU purchase agreements, reported totals). Futures and indices open a third route: exposure to the compute price itself, neither stocks nor bonds. Oil became an asset class bought by pension funds only after gaining a futures market in the 1980s. This is the same process.

⑵ Creates a collateral reference price. GPU-backed loans and data-center securitization (ABS: pooling assets and selling bonds) currently rely on appraisals. A public curve standardizes margins, haircuts (the discount applied to collateral) and revaluations, while more credit can attach to the same collateral. This is the index provider's own description, not our interpretation: reference prices make contracts transparent and planning possible (primary-source statement, translated). The creators themselves say the purpose of pricing is to bring in more capital.

⑶ Turns a story into a market price — a double edge. Opaque compute transaction prices help sustain the AI growth story. With a curve, that price enters the books every day. Market participants' skepticism (secondary sources, not our invention) gathers here: standardization mismatches actual workloads (structural basis), hardware replacement ages benchmarks, and finance has never lived through an entire GPU depreciation cycle.§06 follows where this double edge cuts.

If compute commoditization builds channels for capital, it belongs beside the Treasury's contemporaneous actions. The buybacks in #14 are not a one-off; they are the ladder's first rung. The Treasury's possible moves for managing rates form this ladder:

Today's rung is observed, not guessed. Two scenes:

Scene 1 — The announcement effect did not last three days. On the buyback-doubling announcement (8/19), US 30-year yields fell 5.28→5.19%; only real yields dropped, −9bp (#14's fingerprint). Yet at the 8/21 close, the US 30-year was 5.27%, just 1bp from the day before the announcement. Real yields returned to 3.00% too. Only one thing remained: breakeven inflation (BEI) rose from 2.30 to 2.34%. If an apparent suppression of rates leaves only higher inflation expectations three days later, the market may have begun to read intervention as weakness rather than strength. That is why gauge ④ triggered on 8/21.

★One restraint: buyback ammunition has not yet been deployed (starts 9/9). So far, this is the fading of an announcement effect, not proof that intervention failed. September's actual purchases will show whether the fingerprint appears again.

Scene 2 — Shorter maturities are showing up in outstanding debt. At end-July, bills maturing within 1 year accounted for this share of marketable US Treasuries: 22.2%. That is 2.2pp above the Treasury advisory committee's (TBAC) conventional 15–20% band, and +1.5pp over 12 months. The denominator is marketable Treasuries (USD 31.5 trillion), unlike the news's USD 40 trillion total debt, which includes nonmarketable debt. Exceeding the band is not itself a risk, but it raises refinancing frequency and sensitivity to short-term rates by that excess: the bill for rung 1.

Now place the calendars side by side: juxtaposition, not causation.

The compute-futures listing falls in the middle of the buyback window. The window of rate management and the launch of a new asset class fall within the same two months. That calendar joins this report's title and subtitle.

Why climb this ladder, even managing rates? The US's external financing reveals a structure rather than a prescription. Follow foreigners' US securities net buying over 12 months in Treasury International Capital (TIC) data:

Latest 12-month foreign net buying (through 2026-06): equities USD 919.5 billion vs Treasuries USD 294.4 billion — 3.1x. June alone saw USD 181.4 billion in equity net buying, near a monthly record. The main channel absorbing overseas capital is already equities — the growth story rather than Treasuries. Basis note: this holdings-based SLT series differs in level from the previously cited transaction-based figures (USD 780 billion annually, 2.4x). The direction is the same; the ratio is stronger. Do not mix them.

In this structure, moderate rate management plus a weak dollar makes the absorption mechanism stronger. Currency-hedging asymmetry is the mechanism. For foreign investors, expectations of a weaker dollar hurt bonds first: a few percent of currency losses erases a 4–5% yield, while hedging erases the rate differential. Stocks differ: their volatility makes currency risk relatively smaller, and large companies' overseas revenues mean a weaker dollar raises earnings, providing a partial natural hedge. Lower real yields and rising inflation expectations (§03's residue) therefore shift the advantage from fixed nominal claims (bonds) to growth claims (equities). A negative equity risk premium (ERP), −1.29pp, can then be read as a possible repression equilibrium rather than an anomaly, if equities are the only growth claim in a world of managed bond yields. If the growth story breaks, this reading fails: the other side of #13's narrative-dependence reading.

Crypto institutionalization is rung 2's most American move. On top of GENIUS, which requires short-term Treasuries for stablecoin reserves, tokenized Treasuries are growing: Treasury assets packaged as blockchain funds. But each channel has a different bill-conversion ratio:

Measure where money is actually flowing now:

Over the past 30 days, stablecoin supply changed by−USD 2 billion, spot ETFs by+USD 1.8 billion, and tokenized Treasuries by+USD 100 million. Most inflows are entering the channel with bill conversion 0. Tokenized Treasuries total USD 16.6 billion, about 5% of stablecoins' USD 308 billion. Today's assessment is clear:"More crypto means more bill demand" does not hold now. The growing channel happens to be the one that does not buy Treasuries.

Three routes could overturn this assessment. ⑴Interest-seeking money leaves: payment stablecoins cannot pay interest, so money seeking yield must move to tokenized Treasuries or yield-bearing wrappers. Reports say yield-bearing stablecoins exceed half of supply, but no separate tally was obtained; this remains a gap. ⑵Collateral approval: if tokenized Treasuries become exchange or clearinghouse margin collateral, collateral demand becomes bill demand. ⑶ Legislation: the only event that added fuel regardless of prices was legislation (stablecoins +20% while bitcoin −21% during GENIUS). The 9/15 procedural vote is the next test. The indicators are simple: Does stablecoin supply expand again, and does tokenization growth move from hundreds of millions to trillions?

Each rung lowers one price while leaving a bill elsewhere. Start with those already delivered this week.

⑴ The "smell of QE" — the expectations bill. Buybacks are not QE (#14: the money stock stays unchanged; only maturity composition changes). That is true. Yet on 8/21, when bitcoin passed USD 75,000, market commentary attributed the rally to Treasury buybacks: "When markets smell QE, bitcoin moves" (secondary-source quotation, translated). Whatever the facts, reading it as QE can produce QE-like effects through expectations. That reading matches §03's residue (BEI +4bp). Lower real yields plus higher inflation expectations favor real-like assets without cash flows; observed sensitivities support this (bitcoin correlation with real 30-year yields −0.44, with BEI +0.06). Gold's smaller rise under the same combination also fits: it had already priced this in and is surrendering its premium (regression residual 1.4σ→0.9σ).

⑵ The won — exporting the cost.#14 §03's unmanaged plumbing came under pressure again this week. Across 4 currencies, the won's average+11.5% appreciation (from 6/5) continued its trend. A weak dollar fuels the US absorption mechanism, but brings currency appreciation and a revaluation of export profits to allies. Two faces of the same policy: the largest bill from the US's moves lands overseas.

⑶ The US–Japan gap — two countries make a gap.#14 said US suppression would narrow the gap and hasten yen-carry unwinding. The day after publication, it corrected that: the gap widened to 184bp. Japan's 30-year yield fell from a record 4.096% to below 4%, at 3.995%. The reason makes the contrast: demand returned at auction. In the same week, the same long-bond stress eased through an issuer's purchase announcement in the US (given back in three days) and returning market demand in Japan (still holding). A side-by-side experiment in managed and market-set prices.

⑷ Refinancing risk — a future bill. The 22.2% bill share (§03) saves on today's long-term rates but burdens tomorrow's refinancing. When short-term rates rise, that bill arrives all at once: rung 1's structural cost.

⑸ The compute curve — a boomerang. The protagonist of this report. A curve activates §02's third function: marking the story to market. We observe new B200 rentals at 2.06x older H100 rentals, narrower than the performance gap; China–West token prices differ by 5.45x, and Chinese open models have 55.8% share. Dumping pressure already appears in actual transactions. The NYMEX curve will put that pressure on a public ledger every day. The architecture raising capital against a growth story (§04) is opening a market that formally prices that collateral for the first time. Expanding the absorption mechanism also stress-tests the story. It starts on October 5.

This table is the gauge panel above the title. Like #13 and #14, it is a live widget that updates with new observations after publication. Publication values remain in this table. New needles are bill share, the TIC ratio and 3 channels; ERP + cracks returns. Samsung board and SK hynix cancellation needles keep updating live on #14's gauge panel. They have moved slots, not stopped observing.

This edition followed a force trying to manage prices: rates through a ladder, growth through commoditization, crypto through legislation. Three things lie beyond that management: expectations. August's observations show that suppressing prices can instead lift inflation expectations. BEI is not management's target; it is management's report card. Dumping: pricing strategies across the Pacific lie outside NYMEX's rulebook. An open curve brings them inside. And demand itself: Japan's 30-year auction showed that market-set prices last longer than managed ones. In an era of management, unmanaged prices provide the most valuable information.

That leaves two dates to watch: October 5, when compute enters a public price ledger and the story first becomes a market price. Then November 4, when the buyback window closes and the next issuance plan appears. Does management continue, or does the price return? Somewhere between those dates, the next edition begins.