what i build | agent's take | portfolio | agent picks | ai bellwethers | next fortnight | companies i love | gear i'm hunting | house plan | 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 from October 5, 2026 it becomes a CME futures contract, so the market will show a forward curve as well as spot. Right now it is saying the opposite of discounting: the Silicon Data index sits at $2.53/hr, 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 once it lists (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) 2026-09-28 | $2.70/hr index (+3.8% 7d) vs $2.66 Sep 15; H200 $3.33 (+1.8%); RunPod H100 SXM $3.49/$2.69, Lambda $3.99-4.29 unchanged | up | bullish | This is the marginal clearing price of installed AI compute and the most direct public read on whether capacity is being discounted; it is published daily and becomes a CME-listed futures contract on Oct 5, 2026. watch: Bearish flip if the Silicon Data H100 index holds below ~$2.00/hr for a month, if spot trades below the one-year contract rate (inversion), or if the CME H100 curve lists in steep (>15%) backwardation; the Oct-2025 trough of $1.70 is the floor to remember. — source |
| B200 / Blackwell rental price, $/GPU-hour (on-demand) 2026-09-28 | Silicon Data B200 $5.87 (+2.3% 7d) vs $5.72 Sep 15; B300 index $6.78 (-2.4%); RunPod B200 $5.98/$6.79, Lambda B200 $6.69-6.99 unchanged; B200/H100 ~2.17x | up | bullish | Blackwell is where new supply lands, so this is the first place a glut would show; if B200 rates fell toward H100 rates while Rubin ships, capacity would be outrunning demand. watch: Bearish if tier-1 neocloud B200 on-demand drops below ~$4.50/hr, if the B200/H100 price ratio compresses below ~2x, or if lead times for 8-GPU Blackwell nodes fall to same-week availability. — source |
| Legacy-GPU residual value: used H100 prices, A100/H100 rates and booked useful lives 2026-09-28 | A100 index $1.57/hr (0.0% 7d), above the $1.50 trigger; A100-hyperscaler $3.75 flat; used H100 SXM ~$18-27K by age vs $30K list (Jun); MSFT/Meta lives unchanged | flat | bullish | Depreciation is the biggest cost in a GPU cloud; if old chips still earn money the 5-6 year accounting lives hold and no write-downs are coming, whereas collapsing used-H100 prices would be the first hard evidence of oversupply. watch: Bearish if 8-GPU H100 servers trade below ~50% of new-build cost on the secondary market, if A100/H100 rental rates fall below ~$1.50/hr, or if any hyperscaler extends server lives beyond 6 years to flatter earnings; also watch for Hopper impairments at neoclouds. — source |
| All-in AI data-center build cost, $ per GW (with GPUs / ex-GPUs) 2026-09-28 | ~$47B/GW Vera Rubin (≈$32B racks + ≈$15B shell/power/cooling); ~$60B/GW for gas-powered campus incl. generation (Orennia); ~$1.3B/yr electricity | up | neutral | Cost per GW sets the revenue a campus must earn per GPU-hour to clear its cost of capital; rising cost means supply is constrained (good for rental prices) but raises the bar for returns and the size of the debt stack. watch: Discounting signal if Nvidia rack ASPs or $/GW start falling (Nvidia gross margin guided below ~72%, rack prices cut) while rental $/hr does not rise; bubble-risk signal if ex-GPU cost (power, shell, turbines) keeps inflating faster than rental rates. — source |
| Server DRAM / HBM contract price momentum (QoQ) and DRAM spot 2026-09-28 | TrendForce raised its 4Q26 contract outlook (Sep 24), from +3-8% previously; enterprise SSD +23-28% QoQ, NAND +15-20%; DDR5 16Gb spot $57.83 (Sep 28) vs $54.33 (Sep 14); DDR4 $83.66 vs $89.25 | up | bullish | Memory is the tightest physical input to AI servers (HBM is 25-30% of a Rubin rack) and contract prices are negotiated quarterly, so this is the earliest cyclical input to turn; a price rollover would precede GPU discounting. watch: Bearish when server DRAM or HBM contract prices go negative QoQ, when 2027 HBM4 contracts are signed at lower prices, or when DRAMeXchange DDR5 spot falls >20% from peak; also watch whether CSPs start cutting memory per system, which TrendForce flags as their cost response. Near term: Micron's FQ4 print Sep 30 (guide $50B +/-1B, ~86% GM) and its FQ1 FY27 guide; Samsung Q3 prelim Oct 7-8 (consensus ~107T won op profit). — source |
| Power constraint: ERCOT interconnection queue, gas-turbine backlog and lead times 2026-09-28 | 474 GW ERCOT queue (~90% data centers); Batch Zero screened ~250-300 projects / ~200 GW; Texas freeze widened Sep 21 to all TCEQ data-center permits; BNEF: 49.8 GW delayed; GE Vernova 116 GW backlog | up | bullish | Power, not chips, is now the binding constraint on new capacity, which keeps installed GPUs scarce and rental prices firm; it also reveals how much announced demand is speculative once regulators force projects to prove financing. watch: ERCOT files its Batch Zero eligibility-verification and community-impact reports by Dec 10, 2026 for the Dec 17 PUCT open meeting, with RFIs going out Aug-Sep and cure periods Oct-Nov: large queue withdrawals (BNEF flags 49.8 GW at risk) would show phantom demand. Turbine slot cancellations or resales, falling data-center PPA prices, or turbine prices retreating from the $600/kW 2027 forecast would signal the build is slowing. — source |
| Hyperscaler capex vs cloud backlog (RPO) growth, and capex as % of cloud revenue 2026-09-28 | Oracle RPO $664B (+$26B q/q); Goldman now sees US hyperscaler capex $1.4T in 2027 (Sep 21); Anthropic-Akamai $11.6B/7yr, expandable to ~$20B (Sep 24); 2026 capex ~$1.0T vs ~$2.4T backlog | up | neutral | Capex is the supply side and RPO is the contracted demand side; as long as backlog grows faster than capex the build is pre-sold, but capex above 100% of cloud revenue with negative free cash flow means the cycle is now funded by debt and equity and is sensitive to any demand wobble. watch: Bearish if RPO growth decelerates below capex growth for two quarters, if 2027 capex consensus (+57% per Morgan Stanley) gets cut or 'digestion' language appears, or if backlog concentration in a few AI labs (OpenAI, Anthropic) produces a cancellation or renegotiation. — source |
| Neocloud financing: CoreWeave loan/bond spreads, CDS, vendor backstops, customer concentration 2026-09-28 | CoreWeave $4.2B 2.875% converts due 2033 (closed Sep 24, upsized from $3B, conv. $97.85); Lambda $926M TLB at SOFR+300, Baa2 (Sep 22); CoreWeave loans still ~SOFR+550; debt $35.1B at Jun 30 | down | neutral | Neoclouds are the levered marginal builder of GPU capacity; their cost of debt, covenants and dependence on Nvidia backstops show whether lenders still believe rental prices will cover depreciation, and they are the first place a compute-price decline turns into forced selling. watch: Escalation if new deals price above SOFR+600 or 11% yields, CDS above 1,000bp, deals pulled or downsized, Nvidia's rent-back backstop extended to tier-1 neoclouds, or a top customer (CoreWeave's A/B/C are 36%/26%/10% of revenue) renegotiates; improvement if spreads retrace toward the April 9.75% level or if CoreWeave can tap convertible/equity-linked money on Nebius-like terms. — source |
| Nvidia data-center revenue growth, gross margin and customer concentration 2026-09-28 | DC rev $89.0B +117% YoY (Q2 FY27), GM 75.0%, Q3 guide $108B at 74%; one direct customer >=10% (16%); five customers = 70% of receivables; supply commitments $279B; OpenAI Ohio backstop cut to <$120B | up | bullish | Nvidia's margin is the purest read on whether GPUs are being discounted at the source, and its concentration plus growing vendor-financing (backstops, campus guarantees) shows how much of demand is circular. watch: Bearish if gross margin is guided below ~72% for more than one quarter, data-center revenue grows <5% sequentially, three-customer concentration exceeds 60% in the 10-Q, DSO climbs past 60 days, or backstop commitments expand beyond the $108.5B disclosed; Q3 FY27 report due ~Nov 2026. — 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 2026-09-28 21:35 UTC.