Between mid‑March and mid‑April, the listed B200 SXM spot series (GPUM.B200.SXM.SPOT) fell from roughly $6.66 to $5.88 per GPU‑hour — an 11.6% decline over 30 days, extending a 28% decline over six months. It is the steepest monthly drop since the series became broad enough for a daily fix to be published.

The move is, at first reading, puzzling. Over the same six‑month window, SemiAnalysis's survey‑based contract index reports H100 1‑year rates up roughly 38% and on‑demand capacity "sold out across all GPU types." Listed neocloud spot and surveyed contract transaction rates have detached. Any decomposition of the listed‑side decline therefore has to reconcile against the contract‑side move, not just against supply fundamentals.

Three explanations fit the observation on the surface:

1. Supply-side. NVIDIA and the major cloud operators ramped B200 allocations faster than expected, pushing inventory into the market and compressing rental rates. Under this reading, B200 is on the front edge of a generational deflation path — the familiar curve that took H100 from scarcity to clearing price over 2024–2025.
2. Demand-side exhaustion. Aggregate demand for B200 training capacity softened over the last 30 days — fewer new contract adds at the biggest operators, more capacity unspoken‑for at the margin, rental rates fall to clear. Under this reading, B200's decline says nothing about Blackwell's competitive lifespan. It is a demand signal, not a product signal.
3. Demand-side absorption. Aggregate demand for B200 has not softened; rather, it has migrated from the listed side of the market into long‑tenor contracts. Tight supply gets allocated to the highest‑conviction buyers first, those buyers lock capacity for one or two years, and list‑facing inventory softens because the marginal GPU‑hour now clears out of sight of the public APIs. Under this reading, B200's decline on gpumarkets.dev is a structural signal about where price discovery sits, not a fundamental signal about chip demand.

The three explanations predict very different futures. If the driver is supply, both listed and contract rates stabilise at lower levels. If the driver is exhaustion, both listed and contract rates fall together. If the driver is absorption, listed rates fall while contract rates rise — which is precisely what has been observed over the six‑month window across both indices.

The estimator is a standard log-linear decomposition with one supply proxy and one demand proxy:

log(p_t) = α + β·log(S_t) + γ·log(D_t) + ε_t

where p_t is the daily B200 SXM spot fix; S_t is a supply proxy — cumulative publicly-announced B200 deployments on a rolling quarter, assembled from NVIDIA's shipment commentary and the capex disclosures of the four largest hyperscalers; D_t is a demand proxy — CoreWeave's reported order-book dollar volume, updated from the most recent quarterly filing and the subsequent earnings call transcript; and ε_t is the residual.

The regression window is the 120 trading days ending on 2026‑04‑11. Coefficients are estimated by ordinary least squares against the panel; standard errors are Newey-West with a 10-day lag to control for autocorrelation in daily price residuals. Both proxies are normalised to their first observation in the window.

Over the 30-day window in question, β (the supply coefficient) has not moved outside its 120-day confidence interval. γ (the demand coefficient) has moved sharply — roughly 1.5 standard deviations below its window mean — and the observed residual tracks γ's movement tightly. The decomposition attributes the decline overwhelmingly to the demand term; the supply term contributes less than a tenth of the observed price drop.

Several corroborating signals are consistent with a demand‑side attribution and difficult to reconcile with the supply hypothesis:

• Cross-tier uniformity. If new B200 inventory were landing in the market, one would expect price declines concentrated in the T1 operator cohort (Lambda, the larger neoclouds) where NVIDIA allocations arrive first. The capacity-weighted companion series (GPUM.B200.SXM.SPOT.cw) declined at a rate within sampling band of the equal-weight headline, arguing against tier-concentrated pressure.
• No upward revision to supply forward-guidance. NVIDIA's most recent quarterly commentary restated — rather than raised — B200 shipment trajectory. Hyperscaler capex filings for the quarter did not revise upward for Blackwell-class capacity. The supply narrative requires an unobserved surge; none is visible in the public record.
• CoreWeave's order-book deceleration. CoreWeave's last two quarterly disclosures show the rate of new contract adds moving from roughly double-digit quarter-over-quarter expansion to a flatter trajectory. That is the single largest input to γ's drop in the regression.

The supply hypothesis is therefore rejected. The remaining question is which demand hypothesis holds: exhaustion (demand is falling) or absorption (demand is migrating into contracts). That question is answered from outside the regression, by the contract‑side surface:

• Contract-side rates are rising, not falling. SemiAnalysis's H100 1‑year contract index went from ~$1.70/hr in October 2025 to ~$2.35/hr by March 2026 — up roughly 38% over the same window in which listed spot fell 28%. On‑demand availability at the largest surveyed operators is described as "sold out across all GPU types." Under the exhaustion hypothesis, contract rates should fall in sympathy with spot. They have done the opposite.
• Forward-capacity announcements are being absorbed, not cancelled. Market‑wide capacity coming online through August–September 2026 is reported as already booked. That is incompatible with exhaustion: exhausting demand does not pre‑book unbuilt capacity. It is consistent with absorption: the highest‑conviction buyers are locking capacity at the expense of spot‑facing inventory.

Taken together: the decomposition rules out supply; the contract‑side surface rules out exhaustion; what remains is demand‑side absorption. Price discovery for B200 has migrated from the listed side of the market into long‑tenor contracts. The 11.6% listed decline is priced as a structural signal about where that price discovery sits, not as a fundamental signal about waning demand.

The B200 curve's 11.6% listed decline is a demand‑absorption signal. It is not a supply glut, it is not waning demand, and it is not a generational judgment on Blackwell. It is a signal about where the market is pricing, not what the market is pricing.

The absorption reading produces a testable implication in the contract‑listed spread. If contract‑side rates (tracked by SemiAnalysis's H100 1‑year index and by surveyed neocloud contract adds) continue to rise while listed rates continue to fall, absorption is confirmed and the listed‑minus‑transacted spread should widen further. If contract rates reverse downward alongside listed rates, the market has shifted from absorption to genuine demand exhaustion, and both surfaces should decline together. If contract rates stall while listed rates stabilise, price discovery has re‑aligned between the two surfaces — the spread closes from above rather than widening.

Each of the three paths is informative. None of them is consistent with reading B200 as a generational downgrade. The chip's rental curve is being set by where its buyers choose to transact — and right now they are choosing contracts.

External contract-side data cited in this note: SemiAnalysis, "The Great GPU Shortage — Rental Capacity — Launching our H100 1 Year Rental Price Index," April 2, 2026. Available at newsletter.semianalysis.com/p/the-great-gpu-shortage-rental-capacity . The H100 1-year contract-price trajectory (~$1.70/hr in October 2025 rising to ~$2.35/hr by March 2026, a ~38% increase) and the "sold out across all GPU types" characterization are drawn from that piece.

Listed-side figures (the 11.6% / 30d and 28% / 6mo B200 SXM spot declines) are derived directly from public/chart-data.js in this repository and are live-verifiable against the daily fix published at 00:30 UTC.

References to CoreWeave's quarterly disclosures, NVIDIA's shipment commentary, and hyperscaler capex filings describe the directional character of the most recent public earnings materials as of April 2026, without citing specific numeric claims. A future revision may upgrade these to quoted figures as researcher review lands.

All daily fixings used in this note are reproducible from github.com/gpu-markets/gpu-markets under CC BY 4.0. The decomposition's supply and demand proxies, the normalisation procedure, and the ordinary-least-squares regression used to estimate α, β, and γ are documented in the research/b200-curve-decomposition/ directory of the same repository, alongside the specific filing excerpts and shipment-commentary rows feeding S_t and D_t.

This note is research commentary, not investment advice. The author holds no position in any entity referenced here.

← Back to Research