GPU Market Pulse — Week of May 11, 2026

A week defined by three structural inflections converging on a single binding constraint. The data across the five daily briefs from May 11 through May 17 argued for a coherent investment lens grounded in four cross-cutting themes: custom-silicon substitution is now a tape signal, not a slide-deck argument; H100 has fragmented into separate regional markets; Blackwell B200 is in a real demand-pull squeeze that is propagating backwards into H100 depreciation; and power access is the binding constraint underneath all three.

Executive Summary

• Custom silicon substitution is now a tape signal. AWS Trainium spot in us-ohio printed +332% over seven days; Inferentia2 in Stockholm moved +153% in a single trading day. Amazon's $225bn Trainium backlog (per Andy Jassy's commentary) frames a multi-quarter capacity migration that is consuming Neuron-family spot inventory in real time. A meaningful slice of training and inference demand is being absorbed by silicon Nvidia does not supply, and that substitution is now showing up in the prints rather than in guidance language.

• H100 has fragmented into separate regional pricing tiers. US on-demand pricing in Virginia has held a perfectly flat ceiling at ~$8.69/hr for four months — almost certainly anchored by enterprise reserved-contract terms rather than market clearing. By contrast, European H100 has bifurcated into a 467% intra-continent spread (Frankfurt $1.78/hr vs. Madrid $10.11/hr), and Iowa's spot floor collapsed under single-provider monopoly pricing. The same hardware is now trading three different stories simultaneously.

• Blackwell B200 is in a real demand-driven squeeze that is propagating backwards. B200 spot in us-ohio reached $6.05/hr (+136% 30-day) without a corresponding move in on-demand pricing — the textbook signature of capacity-constrained scarcity. The second-order effect: the H200→B200 generation-gap depreciation rate has compressed to 2.96% annualised, holding the entire H100 installed-base curve elevated against fundamentals. The reset, when it comes, will be sharp.

• Power access is the binding constraint underneath all three. Behind-the-meter power, Nevada grid constraints, the crypto-to-AI conversion arc (IREN's $3.4bn Nvidia contract, Hut 8's $9.8bn deal), and CoreWeave's structurally short power position all return to the same conclusion: every supply-side hypothesis in this market ultimately reduces to who owns permitted, grid-connected capacity at scale.

Investment lens this week. The asymmetric set-up sits with operators of already-permitted, grid-connected power capacity at scale — both pure-play (CORZ, IREN, Vistra) and crypto-to-AI converts (Hut 8, TeraWulf, Core Scientific). Equity prices in late-stage data centre developers continue to assume pre-permit timelines will be met; the cost of that assumption being wrong has risen materially this week. Within Nvidia's own ecosystem, the read is that H100 residual values are propped up by a closing window rather than steady-state demand.

The Week's Themes

Theme 1 — AWS Trainium and Inferentia moved from "threat to Nvidia" to live demand signal

For most of the last two years, the AWS custom-silicon story has lived primarily as a slide-deck argument: a multi-year competitive threat to Nvidia, with adoption tracked through Amazon's earnings disclosures rather than market data. That changed this week. For the first time, the spot tape for Neuron-family instances showed the substitution narrative materialising in real time across multiple regions and instance classes.

The dominant move was Trainium spot in us-ohio, which printed +332% over seven days — a magnitude that is hard to read as anything other than internal Amazon workload migration outpacing publicly available Neuron capacity. Inferentia2 spot in Stockholm spiked +153% in a single 24-hour window, occurring simultaneously with Nscale's $790m Norway data-centre financing and Denmark's 60 GW grid-queue oversubscription — pointing to a coordinated Nordic capacity-build by AWS rather than incidental tightness. Trainium2 on-demand prices rose 13.1% on the week with no corresponding move in spot, the inverse of the H100 pattern (more on that in Theme 3) and consistent with on-demand catalog repricing rather than spot scarcity.

The interpretation we arrived at by week's end is that spot spikes on Neuron-family instances are not noise. They are real-time demand signals from workloads that have already migrated off Nvidia and now sit on AWS's proprietary silicon. The implication for Nvidia's own spot market is direct: a meaningful slice of training and inference demand is being absorbed by silicon Nvidia does not supply. Until this week, that argument had to be carried entirely on Amazon's own forward statements; now it has tape corroboration.

For cloud buyers: The arbitrage windows on Neuron-family instances are narrowing. Workloads with portable Neuron-compatible code (Hugging Face inference paths, AWS SDK-native training) should accelerate provisioning before the price corrections that typically follow these spikes. Workloads with no Neuron path face higher Nvidia spot prices as substitution removes capacity from the broader pool.

For equity analysts: The Nvidia bear case ("custom ASIC substitution will compress NVDA TAM") has moved from speculative to demonstrated. The size of the impact still depends on multi-quarter migration progress, but the directionality is no longer debatable. AVGO's role as the picks-and-shovels supplier across multiple hyperscaler ASIC programmes (Google TPU, Anthropic, OpenAI, Meta) becomes more important as the substitution arc accelerates.

For data buyers: AWS-region spot data on Neuron instances has now become a leading indicator of broader Nvidia spot pricing, with a lag of approximately 7–14 days based on this week's prints. Watch lists should include TRAINIUM|SPOT|us-ohio, INFERENTIA2|SPOT|eu-north-1, and TRAINIUM2|OD|us-east-2.

Principal risk to this read. Single-event noise in a thin market. Several Neuron-family spot tickers have only one provider quoting prices (AWS itself), which means the prints reflect AWS's own capacity-management choices as much as external demand. A counterfactual reading: AWS is deliberately repricing Neuron spot upward to discourage spot-only workloads and migrate them onto reserved capacity. The Inferentia2 Stockholm signal is the cleanest test — if other Nordic instances do not show similar moves over the next two weeks, the "internal repricing" reading is the more probable one.

Theme 2 — H100 has fragmented into separate regional pricing tiers

Going into this week, the working assumption for H100 had been that pricing was beginning to converge — that the early-2026 dispersion across regions reflected supply-build asymmetries that would normalise over time. The data this week argued the opposite. H100 is no longer trading as a single market; it has fragmented into at least three distinct pricing regimes, each driven by a different structural force.

Tier one — reserved-contract anchored. H100 on-demand in us-virginia held a perfectly flat $8.69/hr ceiling for the fourth consecutive month, with a single brief spike to $11.47/hr in mid-January that snapped back within days. This is not a market-clearing price; it is the marginal terms of large enterprise reserved contracts being passed through as catalog on-demand pricing. Spot in the same region traded 18–37% below the OD level — the gap is the bifurcation premium being captured by buyers who can tolerate spot interruption.

Tier two — single-provider monopoly pricing. Wednesday's brief documented Iowa's H100 spot floor collapsing under what appears to be single-provider concentration. With only one provider quoting H100 spot in the region, the price set is not a market clearing process but a unilateral catalog choice. The collapse — and the precision of the new floor — is the tell.

Tier three — demand-driven regional clearing. European H100 traded the widest intra-continent spread of the year this week: Frankfurt at $1.78/hr versus Madrid at $10.11/hr, a 467% range. Madrid's premium reflects emerging-infrastructure characteristics (limited provider count, regulatory friction); Frankfurt's discount is almost certainly a thin-market subsidised listing — but the broader point holds: within a single continent, the same chip can clear at three orders of magnitude apart.

Across the week, the picture moved from "H100 spot is converging" to "H100 spot is decomposing." Three independent regimes are now visible: reserved-contract anchor in US core regions, monopoly pricing in supply-thin US secondary regions, and demand-driven clearing in Europe. The signal-to-noise ratio of any single H100 spot quote has dropped materially as a result.

For cloud buyers: The arbitrage opportunity on H100 spot is wider than it has been since Q4 2025. Workload-portable users (training jobs that can checkpoint across regions, inference workloads with multi-region failover) should be routing to Frankfurt or Iowa where spot floors are at historical lows. The window is unlikely to last more than 2–4 weeks before AWS rebalances inventory.

For equity analysts: H100 single-price metrics — which are still being published by sell-side analysts as a market summary — are increasingly meaningless. Per-region prints are the only useful signal now. The downward H100 depreciation curve being reported in some equity research notes is a methodology artefact of mixing reserved-contract data with spot data; the true depreciation, on an apples-to-apples basis, is materially flatter than headlines suggest.

For data buyers: The data product opportunity here is in the spread, not the level. Per-region spot dispersion is now an investable signal in itself — particularly as it interacts with the AWS Neuron substitution story (Theme 1) and the B200 supply story (Theme 3).

Principal risk to this read. Reading too much into thin-market prints. Iowa, Stockholm, Madrid, and Frankfurt are all small-provider-count regions; a single capacity change can swing the price meaningfully. Treat Frankfurt's $1.78/hr as a directional signal rather than a tradable rate until it persists for at least two weeks.

Theme 3 — Blackwell B200 is in a real demand-driven squeeze, and it is distorting the H100 curve

By the weekend briefs, the B200 picture had become unambiguous. The hardware is in real supply-constrained scarcity — but the more important finding for investors is the second-order effect: the absence of B200 substitution is artificially propping up the entire installed-base depreciation curve.

B200 spot in us-ohio reached $6.05/hr (+136% over 30 days), corroborated by Oregon at $4.85/hr (+84% 30d). In both regions, on-demand prices remained flat over the same window. This combination — rising spot, flat on-demand — is the textbook signature of demand-pull scarcity in a capacity-constrained pool. A supply-driven move would lift both. The 30-day signal is corroborated by IREN's $3.4bn Nvidia deal (announced May 17), which deploys air-cooled Blackwell at 60 MW in Childress, Texas — but with deployment dated early 2027, not this quarter. The supply ramp is happening; it is not yet here.

The second-order signal emerged on Saturday and Sunday. The H200→B200 generation-gap depreciation rate has compressed to 2.96% annualised, implying an implausible ~34-year useful life. Under normal generation transitions, this rate runs 15–25%: V100→A100 was 22%, A100→H100 was 18%. The H200→B200 figure is roughly an order of magnitude lower than the historical baseline.

The 2.96% depreciation print is not because B200 is genuinely a 34-year asset. It is because B200 supply has not been sufficient to substitute for H200 — and by extension, H100 — at meaningful scale. The depreciation rate is a residual that picks up substitution; when substitution is suppressed by supply constraints, the residual collapses.

The consequence is that H100 — which on a normal supply trajectory should be depreciating at 12–18% annualised by now — is also being held up by the same dynamic. Reserved-contract holders are seeing their installed base hold value much better than discounted cash flow models would imply. This is an unstable equilibrium.

For cloud buyers: B200 spot is going to remain bid through at least Q3 2026. Reserve capacity now if your workload has Blackwell-tier requirements. Conversely, H100-portable workloads will see meaningfully better entry prices once B200 supply normalises — that window is likely 3–6 months out, not sooner.

For equity analysts: This is the most consequential dynamic in the GPU complex right now and the most asymmetric. Companies whose installed-base economics depend on H100 not depreciating — most of the neocloud names (CRWV, NBIS, RZLV) — are running on borrowed time. When B200 capacity normalises, both curves will reprice sharply downward together. The first sustained B200 on-demand listings (not just spot quotes) will be the leading indicator that the window is closing.

For data buyers: The per-region B200 spot prints are now the highest-information data in the GPU complex. Watch B200|SPOT|us-ohio, B200|SPOT|us-oregon, B200|SPOT|us-east-1. Convergence across these three is the signal that normalisation is underway.

Principal risk to this read. Reading the Ohio/Oregon prints as broad signals when they are AWS-specific. Azure and GCP have not yet listed B200 at scale; if their build-out timing diverges from AWS materially, the H100 depreciation suppression may persist longer than 3–6 months. The deeper risk is that DeepSeek-style efficiency gains compress training-compute demand structurally — in which case B200 substitution may never reach the H100 installed base because the broader demand pool shrinks.

Theme 4 — Power access is the binding constraint underneath everything else

Across the week's briefs, every supply-side hypothesis eventually reduced to the same question: who owns permitted, grid-connected power at scale. The point was made in different forms by four separate analyses across the week — and the convergence across independent threads is itself the signal.

Monday's brief framed behind-the-meter power as a near-mandatory architecture for new US AI compute build-outs, citing the multi-month grid-interconnection queues across the major ISO regions (PJM, MISO, ERCOT, CAISO). Wednesday's CoreWeave analysis argued the company is "structurally short power" — its discount pricing strategy assumes power costs that are inconsistent with the marginal grid pricing it will face on new capacity. The crypto-to-AI conversion arc — IREN's $3.4bn five-year Nvidia contract, Hut 8's $9.8bn deal, parallel announcements from TeraWulf, Core Scientific, and Crusoe Energy — establishes that already-permitted (formerly mining) capacity is being repriced as AI-ready infrastructure rather than incremental developments. Saturday's Nevada-grid brief made the constraint geographic: Las Vegas-area data centre power is now structurally limited, forcing compute migration west toward Arizona, New Mexico, and the broader Mountain West.

Four independent analyses, four independent data points, one convergent conclusion. Power is no longer one of several supply-side considerations — it is the binding constraint. The next 18 months of US AI compute build-out will be determined by who has permitted, interconnect-cleared capacity ready to host racks. Everything downstream (GPU pricing, regional dispersion, hyperscaler capacity decisions) is increasingly downstream of power availability.

For cloud buyers: Multi-region workload portability is now a power-hedging strategy as much as a cost-optimisation strategy. Lock in capacity in regions with confirmed power headroom; deprioritise regions where data centre announcements have outrun interconnect-queue progress (notably parts of Virginia, North Texas, and Nevada).

For equity analysts: The investment lens for AI infrastructure now sorts cleanly into two buckets. Long bias: operators of already-permitted, grid-connected capacity at scale — including pure-play power infrastructure (Vistra, Constellation, NRG), repurposed crypto-to-AI assets (CORZ, IREN, Hut 8, TeraWulf, Core Scientific), and select pipeline operators (Williams, Kinder Morgan) with gas-to-power exposure. Short or underweight bias: late-stage data centre developers whose equity prices in pre-permit timelines being met. The cost of that assumption being wrong is materially higher now than it was at the start of the year.

For data buyers: Regional power-cost data, interconnect-queue progress, and crypto-to-AI conversion announcements have all become first-order signals for AI infrastructure analysis. The pure-pricing-tape view of the market increasingly misses the constraint that drives it. Cross-referencing power-grid data with GPU spot data is the highest-value composite signal currently available.

Principal risk to this read. Federal pre-emption of state grid rules. A federal AI-permitting framework that overrides state-level interconnect queues would reset the asymmetric set-up in favour of late-stage developers. The probability is non-trivial but not high in the next 12 months; the political economy still favours states retaining grid authority. A secondary risk is demand destruction (the DeepSeek scenario) reducing the binding constraint pressure without altering the supply side.

Connections Between Themes

The four themes are not independent. They reinforce each other in ways that compound the investment implications.

The most direct connection runs between Themes 1 and 3. AWS's Neuron family is absorbing inference and training demand that would otherwise flow into the Nvidia spot pool. That substitution slows the rate at which Nvidia installed-base depreciation should normalise — because some of the demand that would have rolled forward to B200 is now exiting the Nvidia complex entirely. The depreciation curve flattening (Theme 3) is partially a consequence of the substitution signal (Theme 1), not just a B200 supply problem.

Themes 2 and 3 connect through a regional channel. Frankfurt's H100 $1.78/hr discount looks less anomalous when read in the context of B200 supply being concentrated in Ohio and Oregon. European regions don't have B200 to substitute toward, so European H100 spot can hold demand-driven price levels without depreciating against an alternative. Madrid at $10.11/hr is the same thing in the other direction — a region where supply is scarce, no alternative exists, and prices reflect actual demand pressure.

Theme 4 sits underneath all three. The reason B200 supply is constrained in some regions and abundant in others is power and grid access. The reason Neuron family expansion is happening preferentially in Nordic regions is power and grid access. The reason H100 European pricing is bifurcated is the underlying differential in regional power costs feeding through into provider economics.

The investment lens that ties these together: position toward operators with permitted power capacity (Theme 4), benefit from the H100 depreciation window before it closes (Theme 3), avoid neocloud names exposed to the depreciation reset (Theme 3 risk), and treat AWS Neuron family spot prices as a leading indicator for broader Nvidia spot direction (Theme 1).

Featured Data Cuts

Two views of the underlying pricing tape that are not in the daily briefs, drawn from Signwl's regional and instance-class data.

B200 spot dispersion across AWS regions (30-day prints)

Read: B200 spot is not yet a global market. Three of five major AWS regions have no public quote; the three that do are pricing the same hardware in a 119% range. Cross-region convergence is the leading indicator for supply normalisation and the close of the H100 depreciation window.

AWS Neuron family — 7-day spot momentum

Read: Neuron-family spot is now bifurcating across regions in a pattern that resembles H100 fragmentation (Theme 2) but on a steeper time-scale. The cross-region rotation — particularly Ohio +332% vs. Oregon −51% on similar Trainium instances — looks like internal AWS capacity rebalancing rather than uniform demand growth. Treat region-specific reads accordingly.

What to Watch Next Week

The themes set up four specific watch points for the week ahead, each with confirm/deny criteria so this brief sets up next week's read.

1. B200 cross-region convergence. Watch Ohio and Oregon spot prints against us-east-1. If Ohio and Oregon stay elevated (>$5/hr) while us-east-1 stays soft (<$3/hr), the supply imbalance is regional rather than absolute and the arbitrage window persists. Convergence (all three in the $3.50–$4.50 range) would signal AWS has rebalanced inventory; the H100 depreciation window starts to close at that point.

2. Inferentia2 European expansion. Stockholm's single-day spike likely reflects new capacity coming online. The Nordic capacity-build thesis confirms if Dublin or Frankfurt show similar +50% intraday moves; denies if both stay flat while Stockholm corrects. The Nscale $790m financing and the Denmark grid-queue oversubscription both point to a coordinated expansion; the prints should follow.

3. CoreWeave Q2 disclosure (late July). Not next week, but the watch starts now. The margin-trap thesis becomes investable only if CRWV's next guide surfaces the gap between its current discount pricing and the marginal power cost it will face on new capacity. Watch any pre-guide commentary at conferences, deal announcements, or analyst day events that hint at margin pressure.

4. First sustained B200 on-demand listings. B200 today is largely a spot-only quote across AWS, Azure, and GCP. The appearance of stable on-demand pricing (priced over 30+ consecutive days) will mark the start of supply normalisation. This is the leading indicator for the H100 depreciation reset described in Theme 3.

Calendar items worth flagging. Earnings from Vistra (May 22), Constellation Energy (Jun 1), and Equinix (week of May 28) will all carry implications for the power-constraint thesis. Any updates to ERCOT or PJM interconnect-queue rules in the same window would be material. The European Commission is also expected to clarify guidance on the EU AI Act's data-centre provisions in early June; that release will reset the European H100 pricing reads in Theme 2.

Market Snapshot

Blended GPU compute costs rose 3.1% this week across major cloud providers, led by ALVEO_U30's 58.4% gain.

Top Movers

Blended pricing = average of spot, on-demand, and 1-year reserved rates across major cloud providers.

Training vs Inference

Training-class GPU pricing held steady this week (avg $4.29/hr, +0.1% WoW), while inference-class pricing rose (avg $0.53/hr, +5.8% WoW).

The training-to-inference price ratio stands at 8.1x — narrowing compared to last week. The elevated spread suggests strong demand for training compute relative to inference, consistent with ongoing large model training activity.

Regional Spotlight: South America

South America trades at a 31% premium to global averages this week, with 19 GPU types available across the region. The most expensive GPUs in the region are GB200 ($23.02/hr), H200 ($11.49/hr), H100 ($9.05/hr). The 31% regional premium reflects emerging infrastructure and limited provider competition.

For detailed pricing data across all South America sub-regions, see the full regional profile.

Implications

For cloud buyers: Asia Pacific continues to offer the lowest average GPU pricing ($2.13/hr blended average). For workloads with regional flexibility, the gap between Asia Pacific and Middle East is $1.97/hr — a 92% premium. Compare regional pricing →

For semiconductor analysts: GPU pricing trends remain broadly stable this week. H100 (−0.4% WoW) and MI300X (+0.0% WoW) are tracking within normal ranges. Blackwell (B200) blended pricing at $8.01/hr (+0.3% WoW) provides an early read on next-generation adoption curves. View all GPU profiles →

For GPU investors: Rising rental rates are improving returns for existing GPU deployments. Model scenarios with the GPU ROI Calculator →

Things to Watch Out For in the Next Weekly Brief

Next Tuesday's edition will track the four themes forward, with particular attention to the data points most likely to move conviction:

• B200 cross-region convergence — refreshed dispersion table against the next seven days of prints. If Ohio and Oregon stay >$5/hr while us-east-1 stays <$3/hr, the regional supply imbalance read holds and Theme 3 remains the most actionable view in the complex.
• Inferentia2 European spread — whether the Stockholm spike propagates to Dublin and Frankfurt. A coordinated Nordic build confirms Theme 1; a Stockholm-only correction softens it.
• Vistra Q1 (May 22) and Constellation Energy commentary (Jun 1) — both calls will reset the power-constraint read for Theme 4. We will focus on AI-driven demand disclosures and any commentary on grid-interconnect timeline progress.
• First weekly H100 depreciation snapshot — comparing the current implied rate against the V100→A100 and A100→H100 historical baselines. The current 2.96% print is the anomaly; we will track when it begins to normalise.
• Computex pre-positioning — any commentary or partner announcements ahead of the Computex window that recalibrates the B200 supply outlook.

If there are specific tickers, regions, or instance classes you would like covered in the next brief, reply to the newsletter — we incorporate reader requests directly into the following week's analysis.