what i build | agent's take | portfolio | agent picks | ai bellwethers | next fortnight | companies i love | gear i'm hunting | news | research log
The single best AI-bubble bellwether is the rental price of an installed Nvidia GPU, measured in dollars per GPU-hour, and specifically the H100: it is the daily marginal clearing price for AI compute, it is published by independent indexes (Silicon Data's SDH100RT, SemiAnalysis's one-year contract index), and a CME futures contract on it is pending: the CFTC extended its review to November 9, 2026, so a forward curve is coming but not yet listed. Right now it is saying the opposite of discounting: the Silicon Data index sits at $2.75/hr (Sep 30), a new high, one-year contract rates rose about 40% from $1.70 to $2.35 between October 2025 and March 2026, spot has run at roughly twice contract, and a four-year-old chip is renting for more than it did a year ago while CoreWeave says it has contracted 2020-vintage GPUs out to 2029 at full price. Pair it with three things: neocloud credit spreads (CoreWeave's July loan cleared at SOFR+550 for a 10.44% yield with maintenance covenants, and its CDS implies roughly a coin-flip default probability), which is the financing leg and the one signal already flashing red; server DRAM and HBM contract-price momentum (+13-18% QoQ in 3Q26 after +93-98% in 1Q26, with 4Q26 guided to only +3-8%), because memory is the first physical input that will roll over; and hyperscaler backlog growth versus capex growth (Microsoft commercial RPO +84%, Oracle RPO +363%, roughly $2.4 trillion of committed cloud backlog against about $1 trillion of 2026 capex), because the bubble only deflates if the buyers stop signing. Over the next one to two quarters watch Nvidia's August 26 print for gross margin and customer concentration (three direct customers were already 54% of revenue), the shape of the CME H100/B200 curve if the CFTC clears it in November (a steep backwardation would be the first public forecast of discounting), whether 4Q26 memory contract talks and 2027 HBM4 pricing turn negative, and what ERCOT's December audit says about how much of Texas's 474 GW interconnection queue is phantom demand. The threshold that would flip the call is simple: the H100 index sustained below about $2.00/hr, or spot trading below the one-year contract rate, while Blackwell and Rubin volumes keep arriving; until then compute is being sold at a premium, not a discount, and the risk lives in who is borrowing to build it rather than in the price of the product.
| indicator | reading | trend | signal | why it matters / what to watch |
|---|---|---|---|---|
| H100 rental price, $/GPU-hour (Silicon Data index; 1-yr contract and spot) 2026-10-18 | Silicon Data SDH100RT $2.82/hr as of Oct 17, down 1c on the day, +6.4% 7d, 3c under the $2.85 high. H200 $3.51 (+6.0% 7d), up 6c; A100 $1.53 (+3.2%); MI300X $2.64 (0.0%). CME H100/B200 futures: no listing news found. | up | bullish | Hopper spot firmed through the Oct 8 OpenAI revenue scare: every Silicon Data index rose over the week, so rented compute is still being bid, not discounted. watch: Bearish flip if SDH100RT prints below ~$2.00 for a month or spot drops under the 1-yr contract rate. Next: whether the CoreWeave credit scare keeps pulling Hopper rents down, and the CME/CFTC listing decision. — source |
| B200 rental price, $/GPU-hour (Silicon Data index) and B200/H100 ratio 2026-10-18 | Silicon Data B200 $6.12 as of Oct 17 (+7.0% 7d), up 11c to a new high; B300 $7.01, up 5c. B200/H100 ratio 2.17x (6.12/2.82), up from 2.12x. Both above the $4.50 and 2x triggers. | up | bullish | Blackwell rents rose again and B300 rose faster than B200; the newest capacity is still clearing at a premium, which is what a supply-constrained market looks like. watch: Bearish if B200 drops below $4.50 or the B200/H100 ratio falls under 2x. Also watch whether B300 keeps widening its premium over B200 as Rubin nears. — source |
| Hopper residual value: A100 index floor, used H100 prices, GPU useful-life accounting 2026-10-18 | A100 index $1.53 Oct 17 (+3.2% 7d; trigger $1.50), up 1c, three cents above the floor. H200 $3.51, up 6c. Silicon Data residual study (Jan 30 data): used H100 ~65-75% of new at 2 yrs, 45-55% at 3 yrs; low end $12-16K. No new impairments found. | flat | bullish | A100 is still a cent or three above the $1.50 floor and ticked up this week; the residual-value question is unresolved rather than deteriorating. watch: A100 index under $1.50; any hyperscaler or neocloud shortening GPU useful life or booking a Hopper impairment in Q3 results (late Oct onward); used H100 SXM asks drifting below $18K. — source |
| Data-centre cost per GW and the cost of capital (10-yr Treasury) 2026-10-18 | 10-yr Treasury par yield 5.09% on Oct 16 (Treasury, final), down 1bp, a new low for the month and 22bp under the Oct 5 high; 2-yr 4.74%, down 3bp. No new campus $/GW prints (last: Vera Rubin ~$47B/GW, gas ~$60B/GW, Jupiter ~$67B/GW). | flat | neutral | Build cost per GW is unchanged and the 10-yr has eased about 20bp from its October high but is still above 5%, so every GW is funded at the most expensive rates in two decades. watch: 10-yr back above 5.30%; any new campus announcement implying more than ~$60B/GW; Amazon's newly disclosed data-centre deal terms now that it is dropping NDAs. — source |
| Memory prices: DRAM spot and contract, HBM, Samsung and Micron readthrough 2026-10-18 | TrendForce spot Oct 16: DDR5 16Gb $60.60 (+0.28%), a high; DDR4 16Gb $83.66 (flat); DDR4 8Gb $45.97 (flat). No new session since. Public contract table still 2H Aug; no Q4 DRAM or HBM contract figures found. | up | bullish | Server-grade DDR5 spot keeps setting highs through the CXMT selloff; the only soft spot is DDR4 and PC demand, which does not touch server DRAM or HBM. watch: TrendForce's September/October contract update (a Sept 30 update exists but is paywalled); the first DDR5 16Gb spot decline as CXMT's 4F2 server DDR5 nears launch; DDR4 16Gb spot rolling over from $83-84; Samsung full Q3 late Oct. — source |
| Power bottleneck: ERCOT/PJM queues, Texas permit freeze, turbine backlog 2026-10-18 | FERC (Oct 9 order) rejected PJM's interim large-load service, which let data centres connect without new capacity if curtailable, and its bring-your-own-capacity rules; approved the large-load definition and registry. ERCOT 474 GW large-load queue; TCEQ freeze since Sep 21, report due Oct 19 (not yet public); TWDB report due Oct 14 not yet found. | flat | bullish | Grid capacity is still the binding constraint on how fast compute can be deployed: PJM is issuing tight-conditions advisories in October and only just fast-tracking its first new generators, which keeps existing GPU capacity scarce. watch: TCEQ compliance report Oct 19; the overdue TWDB data-centre report; PJM's revised large-load filing after FERC's Oct 9 rejection; ERCOT Batch Zero timing; GE Vernova Q3 backlog (late Oct). — source |
| Capex vs backlog: hyperscaler spend, RPO, AI-lab commitments and buyer revenue quality 2026-10-18 | TSMC call (Oct 15): 2026 capex kept at $60-64B; 2026 revenue growth raised to about 44% in USD; 2026-28 capex 'significantly higher' than the prior three years; AI-accelerator CAGR (mid-to-high 50s%, 2024-29) 'probably will be higher'; demand above supply, 'no sign of slowdown'. Q4 guide $52.3-54.3B. | down | neutral | Equipment and foundry demand say the build keeps accelerating, but the biggest buyer's revenue base is ~30% smaller than the market assumed, so backlog coverage by customer revenue got worse on paper; it is an accounting gap, not a demand drop, which keeps this neutral rather than bearish. watch: OpenAI's $30B round at $1.4T pre-money closing or re-priced; Meta Oct 22 AEDT and other Q3 hyperscaler capex guides late Oct; any capex cut or delay tied to neocloud credit stress. — source |
| Neocloud credit: CoreWeave leverage and bond yields, neocloud and chip financing 2026-10-18 | CoreWeave fell more than 11% Friday, nearly 25% on the week; its Oct 2031 bond briefly yielded over 16% Thursday and lost about 3c this week, a fifth weekly fall (Reuters). 5-yr CDS record intraday 1,194.5bp (ICE). 2032 notes below 78c (Bloomberg via Korea Herald). Nebius -12%, Applied Digital and IREN -11% Friday. | down | bearish | CoreWeave's unsecured bonds now price a distressed borrower: the financing leg of the AI build is breaking first, even while GPU rents hold. watch: CoreWeave 2032 bond back above 85c or below 75c; any DDTL draw, convert or equity raise; Q3 results early Nov; Nebius and Oracle debt pricing; OpenAI's $30B round closing; Lambda, Crusoe raises. — source |
| Nvidia data centre: revenue, margin, concentration, backstops and demand signals 2026-10-18 | NVDA -1.4% Friday, -6.6% on the week. Huang in Berlin: AI selloff 'going to pass'. GTC Berlin Oct 20-22, Huang keynote Oct 21 (Tempodrom). German industrial-AI partnership expected next week (Reuters/Handelsblatt). DC rev $89.0B (+117%), GM 75.0%, Q3 guide $108B. | up | bullish | Fundamentals and guidance are unchanged and TSMC's HPC growth confirms volumes, but buyer concentration and neocloud credit are now the live risks behind Nvidia's largest customer programmes. watch: GTC Berlin keynote Oct 21; Meta Oct 22 AEDT; any OpenAI round re-pricing; neocloud order deferrals; GTC Washington Nov 30-Dec 3; Q3 FY27 results in late Nov (inference from prior cadence); CFTC/CME GPU futures listing. — source |
signal is from the AI-infrastructure investor's view: bullish = still supply-constrained, bearish = compute being discounted / oversupplied.
Nothing here is financial advice. Holdings and opinions are my own; the "agent" sections are written by an automated research skill and may be wrong. Built 2026-10-18 22:56 UTC.