On April 13, 2026, the Wall Street Journal set Silicon Valley boardrooms on edge with a single data point: the hourly cloud rental for an Nvidia Blackwell GPU had hit $4.08, a 48% jump from $2.75 just two months earlier. The number, sourced from the newly Bloomberg-listed Ornn Compute Price Index, instantly became the most-quoted figure in AI infrastructure, vindicating what neocloud CEOs had been whispering since CES: the industry is staring down its worst compute shortage in five years, and Blackwell capacity is the new oil.
The surge crystallizes a stress test on the entire $650 billion AI capex cycle. Vultr CEO J.J. Kardwell called the squeeze “the most severe compute shortage I’ve seen in over five years of running this company.” CoreWeave has hiked spot rates more than 20% since December 2025 and is forcing small and mid-tier tenants into three-year commitments. Anthropic’s Claude API uptime slipped below industry norms in Q1, OpenAI quietly shuttered the consumer Sora video generator to claw back tokens, and Bank of America’s chip team now projects the GPU supply-demand gap will not close until at least 2029. Wall Street is treating $4.08 as the new floor.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
The $4.08 Print: How the Ornn Index Became the New AI Benchmark
The Ornn Compute Price Index, a spot-market tracker for cloud GPU rentals, was integrated into the Bloomberg Terminal in the first week of April 2026 –, and within five trading sessions it became the most-watched alt-data feed for AI infrastructure analysts. The index’s headline reading on April 13 was a clean $4.08 per Blackwell GPU-hour, capturing spot transactions across more than 30 neocloud providers and arbitraged hyperscaler reserved-capacity sublease markets. Two months earlier, the same index sat at $2.75. The 48% climb was not a quiet drift; it was a step-function repricing that began the second week of March and accelerated through Easter.
Crucially, the Ornn reading reconciles with Silicon Data’s parallel SDB200RT index, which clocked the B200 spot rate at 5.48 on March 30 (peaking at 6.11 on March 25, before settling into the $4.00-$4.10 band that the WSJ amplified). Silicon Data’s daily volatility logs show the B200 rental dropped 1.2% in February before exploding 23.6% in March alone – almost the entire 48% headline number was printed in a four-week window driven by agentic-AI inference contracts, post-training runs from second-tier model labs, and unscheduled inference burn from Anthropic’s Claude 4.6 family and OpenAI’s GPT-5 reasoning rollouts. The reconciliation held into the following quarter: Presenc.ai’s own tracker estimated single B200 rentals at $4.50-$7.00 per hour in Q2 2026, a band that sits above the Ornn print and suggests the April headline, if anything, understated the true ceiling enterprise buyers were paying for guaranteed Blackwell capacity.
What makes the print uncomfortable for hyperscalers is the contrast with last-generation pricing. H100 hyperscaler rentals at AWS, Azure, and GCP have stayed remarkably flat at $7.43-$7.52 per hour on three-year reserved commitments, while neocloud H100 spot has compressed to $2.43-$2.63. The premium customers are now paying for Blackwell – even with FP4 throughput roughly 2.5x a H100 – implies they are willing to lock in any Blackwell hour they can find regardless of margin, because the queue for new HGX B200 cabinets is now measured in calendar quarters, not weeks. Even the workstation-class RTX PRO 6000 Blackwell SKU, marketed as a cheaper on-ramp to the architecture, hasn’t been spared: by August 2026, ThunderCompute was quoting rentals swinging anywhere from $1.11 to $4.50 per hour, while Spheron listed dedicated RTX PRO 6000 Blackwell capacity at $2.29 per hour and spot at just $1.19 per hour – a spread wide enough to show that pricing chaos has spread well beyond the flagship B200 and GB200 parts.
Why Blackwell Capacity Is the New Bottleneck
Three forces converged to break the previous equilibrium. The first is agentic inference. Anthropic disclosed in its Q1 developer update that long-running agent traces – Claude Code sessions, computer-use agents, and Project Glasswing security workloads – burn between 8x and 40x more output tokens than chatbot interactions. OpenAI confirmed at the same time that API token volumes had more than doubled year-over-year. Inference, not training, is now the marginal buyer of Blackwell hours, and unlike training runs which can be paused, agentic workloads are SLA-bound and price-inelastic.
The second is post-training and synthetic-data pipelines. With model labs running multi-week reinforcement-learning sweeps on agent benchmarks (SWE-bench, OSWorld, ARC-AGI-2), demand for elastic Blackwell clusters in the 256-1,024 GPU range has surged. These customers used to buy preemptible H100 capacity; in 2026 they pay neocloud spot for Blackwell because the wall-clock speedup pays for itself in researcher productivity. Lambda, Crusoe, Nebius, and FluidStack have all reported sold-out Blackwell pods for Q2 and Q3 delivery.
The third is the supply side. TSMC’s CoWoS-L advanced packaging – required for the GB200 NVL72 rack-scale system and HGX B200 boards – remains the choke point. TSMC’s Q1 2026 capex of $56 billion is partly aimed at doubling CoWoS capacity to a 75,000-wafer monthly equivalent by year-end, but the new lines do not light up until late Q3. HBM3e memory pricing from SK Hynix, Samsung, and Micron rose roughly 20% for 2026 delivery contracts, and Nvidia passed those increases through to its MSRP in late February. Cloud rental rates absorbed the pass-through with the usual four-to-six-week lag – landing, almost on cue, in the Ornn index.
The Blackwell Rental Repricing in Numbers
| Date | Blackwell Spot ($/GPU-hr) | Ornn Index | 2-Month Change | Key Driver |
|---|---|---|---|---|
| Jan 2, 2026 | $4.40 | 4.40 | – | Holiday inference burn |
| Jan 30, 2026 | $4.29 | 4.29 | -2.5% | Post-holiday softness |
| Feb 13, 2026 | $2.75 | 2.75 | -37.5% | Brief overcapacity dip |
| Mar 25, 2026 | $6.11 | 6.11 | +122% | Intraday peak (agent demand) |
| Mar 30, 2026 | $5.48 | 5.48 | +99% | Silicon Data SDB200RT close |
| Apr 13, 2026 | $4.08 | 4.08 | +48% | Ornn/WSJ headline print |
The February $2.75 reading is the anchor. It reflects a brief mid-quarter overhang as several training cohorts wound down and CoreWeave released contractual capacity from a delayed Microsoft tenancy. That window slammed shut in early March when both OpenAI and Anthropic upsized inference reservations to support, respectively, the GPT-5 reasoning rollout and the rapid scale-out of Claude Code in the enterprise. By the time the WSJ printed on April 13, every neocloud floor manager was rationing.
CoreWeave, Lambda, Crusoe: Neocloud Winners and Their Pricing Power
For the neocloud cohort, $4.08 is gross-margin gold. CoreWeave, with roughly 470,000 Nvidia GPUs deployed and a backlog north of $35 billion as of its Q4 2025 earnings, telegraphed the price action by quietly amending its standard terms-of-service in late December. The amendment imposed minimum 36-month commitments on new customers below $5 million ARR – a deliberate move to convert spot demand into multi-year deferred revenue. The strategy has paid off: every dollar of incremental Blackwell pricing flows almost directly to operating cash flow, because CoreWeave’s underlying lease payments to Nvidia and Switch are fixed.
Lambda Labs and Crusoe Energy occupy the secondary tier. Lambda has leaned into developer-friendly hourly billing and has been raising headline rates in lockstep with the index, while Crusoe has rerouted incremental Blackwell pods toward its Abilene and Stargate-adjacent campuses to backfill the OpenAI compute pipeline. Nebius, the European spinout of the former Yandex compute business, lifted its Blackwell list price 35% on April 1 and is now sold out through Q3. FluidStack, freshly valued at $18 billion and embedded in Anthropic’s $50 billion compute deal, is effectively unable to take new spot customers at all.
“This isn’t a cycle, it’s a regime change,” argues Stacy Rasgon, senior semiconductor analyst at Bernstein. “When a $4 print sticks for two quarters, it’s not a print, it’s a benchmark. Neoclouds will use it to refinance debt at lower rates; hyperscalers will use it to justify another leg of capex, and customers will use it to renegotiate up.” Rasgon’s note to clients on April 14 quantified the implied EBITDA uplift for CoreWeave at roughly $1.2 billion on an annualized basis if the $4 floor holds through year-end.
Hyperscaler Response: Stable Pricing, Tighter Allocation
The hyperscalers – AWS, Microsoft Azure, Google Cloud – are pursuing a deliberately different playbook. Public Blackwell list prices on the on-demand storefronts have barely budged. AWS p6-b200.48xlarge instances remain advertised at the same hourly rate set at launch; Azure’s ND-B200 v6 series and GCP’s a4-highgpu-8g pricing are similarly anchored. The reason is regulatory optics and customer durability: hyperscaler CFOs cannot afford to be seen surge-pricing strategic AI capacity into Fortune 500 accounts that signed seven-figure annual commitments expecting stable economics.
What is changing is allocation. Both AWS and Azure have moved Blackwell rationing into priority tiers, reserving Enterprise Agreement customers and existing Anthropic/OpenAI workloads ahead of self-service demand. Smaller customers face quota throttles and capacity-not-available errors that did not exist six months ago. In private, hyperscaler account teams are pushing reserved-instance and savings-plan conversions hard – effectively repricing the spot-on-demand gap into longer commitment terms rather than raw hourly rates. “The list price is the press release. The contract is the trade,” one Azure field CTO told Tech Insider on background.
Google Cloud has hedged by re-routing high-volume training customers onto TPU v6e Trillium and TPU v7p (the 121-exaflop “8t” and “8i” generation announced earlier in 2026), reducing pressure on its Blackwell allocation. Microsoft’s response has been the most aggressive on capex: Satya Nadella reiterated on the FY26 Q3 call that Azure would spend “north of $80 billion” on AI infrastructure in calendar 2026, of which a meaningful share targets Blackwell racks already paid for but awaiting power and CoWoS-bound HBM.
Customer Pain: OpenAI Sora, Anthropic SLAs, and the API Token Crunch
The downstream pain is visible in product cuts. OpenAI shuttered consumer Sora in late March, redirecting the underlying inference capacity to the GPT-5 and o-series API. Anthropic, despite its $40 billion Google TPU bet and $50 billion FluidStack partnership, missed its Claude API remained largely operational during March 2026 outages (e.g., March 2-3 and March 11) with only intermittent degraded performance, while web/mobile interfaces failed; Q1 uptime was ~99% or lower per status.claude.com (e.g., claude.ai at 98.21% for March), not meeting 99.9% SLA, with customers in Europe and APAC reporting elevated overload errors for weeks at a stretch. Several Fortune 500 enterprise customers began invoking SLA credits, a first for the company.
Inference-heavy startups felt it worst. Cursor, Replit, Factory AI, and Poolside (mid-collapse on its own Texas data-center bet) all raised effective per-seat prices in March and April, citing compute pass-through. Cursor told customers that input-output token costs from underlying foundation models had risen 30-40% on a blended basis, and pushed many seats from unlimited usage to metered tiers. The “AI free tier” era is, for the moment, on hold.
Even Meta, which famously buys rather than rents, is feeling indirect pressure. The company’s $135 billion 2026 capex commitment includes a meaningful sliver of merchant-cloud spend for surge Llama-5 inference; that envelope is now buying fewer GPU-hours than it did at the December 2025 commit. Mark Zuckerberg’s Q4 talking point – that “the worst outcome would be under-investing in AI” – looks prescient. Meta’s MTIA custom silicon program with Broadcom (2nm, ramping 2027) cannot arrive fast enough.
Blackwell vs. The Competition: AMD MI355X, Google TPU, AWS Trainium3
The obvious question: where is the substitution? AMD’s MI355X is shipping, and MI400 is on the 2026 roadmap with 320 billion transistors and HBM4 – explicitly positioned as a Blackwell competitor on inference. Google’s Trillium TPU v6e and v7p remain captive to GCP and a handful of strategic partners (Anthropic above all). AWS Trainium3 just won the Uber workload in April and is being designed to a stated 50% Nvidia discount. So why isn’t the surge being relieved?
The answer is software gravity and developer inertia. CUDA, cuDNN, TensorRT-LLM, NCCL, and the new NIM microservices remain best-in-class for production inference at scale. Porting a Claude- or GPT-class inference stack to ROCm or Trainium SDK is technically feasible but operationally expensive – the kind of decision a CTO makes once a year, not in response to a two-month spot-price spike. Hyperscalers and neoclouds are also signing multi-year purchase commitments with Nvidia that lock in product priority – meaning any short-term substitution would forfeit long-term allocation.
“AMD has the silicon. What it doesn’t have is the production reference-architecture density that Nvidia just shipped with GB200 NVL72,” says Patrick Moorhead of Moor Insights & Strategy. “Every CIO I talk to has run an MI300X pilot. Almost none have moved their primary inference fleet. The MI355 and MI400 are credible – but credibility is not a CUDA kernel.” The implication is that Blackwell pricing remains effectively unhedged for the next 12-18 months.
Spot vs. Reserved: How the Two-Tier Market Now Works
| GPU SKU | Hyperscaler Reserved ($/hr) | Neocloud Spot ($/hr) | Spot Premium / Discount | Typical Tenor |
|---|---|---|---|---|
| Nvidia H100 SXM5 80GB | $7.43 – $7.52 | $2.43 – $2.63 | -66% | 1-3 yr commit vs hourly |
| Nvidia H200 SXM5 141GB | $8.20 – $8.80 | $3.10 – $3.60 | -58% | 1-3 yr commit vs hourly |
| Nvidia B200 HGX 192GB | $10.50 – $11.20 | $4.08 (Ornn spot) | -62% | 3 yr min (neocloud) |
| Nvidia GB200 NVL72 (per GPU) | $12.00 – $13.00 | $5.50 – $6.50 | -50% | 3 yr commit (rack-level) |
| AMD MI300X 192GB | $5.20 – $5.80 | $2.10 – $2.50 | -58% | 1 yr commit vs hourly |
| Google TPU v6e (per chip) | $2.70 – $3.10 | captive / no spot | n/a | GCP only |
The table illustrates the two-tier market. Hyperscaler list prices are stable but high; neocloud spot is volatile but cheaper in absolute terms. The implicit message: customers who locked Blackwell reservations in mid-2025 are sitting on what is now in-the-money compute, while marginal buyers face the Ornn floor. The economic transfer is meaningful – a 10,000-GPU training cluster running 24/7 at a $4 floor is a $350 million annual run-rate, twice the bill at the prior $2.75 print. That discount dynamic extends down the product stack too: as of August 2026, Spheron was still able to offer RTX PRO 6000 Blackwell spot capacity for as little as $1.19 per hour, a reminder that smaller workloads willing to step down from full B200/GB200 clusters can still find meaningfully cheaper Blackwell-class compute outside the datacenter-grade tier.
The Capex Bubble Debate: Is $4.08 a Signal or a Bubble?
The bullish read on the price surge is straightforward: it validates the $650 billion big-tech AI capex cycle. If GPUs are clearing at $4+ per hour on three-year tenors, every dollar of Nvidia, Broadcom, AMD, TSMC, ASML, Marvell, and Micron capex is well-spent and the supply curve is the bottleneck. Goldman Sachs’ semiconductor team upgraded Nvidia’s price target on April 14, citing “structurally higher rental yields supporting customer ROIC.”
The bearish read is uglier: the $4.08 print is exactly what a late-cycle bubble looks like. The thesis runs that AI labs are paying $4+ for inputs that they cannot yet monetize at margin – see OpenAI’s recent reported revenue miss versus the $400 billion Stargate compute commitment –, and that the moment agentic workloads disappoint at scale, demand evaporates, leaving neoclouds holding overpriced, depreciating silicon on three-year leases. The $2 trillion SaaS stock-crash narrative from earlier in 2026 lurks under this view.
“What worries me is that nobody is yet asking whether the agentic demand curve is real or marketing,” says Gil Luria, head of technology research at D.A. Davidson. “If Claude Code and Cursor genuinely 5x developer productivity, the capex is cheap. If the productivity gains are 20-30% and concentrated in a thin slice of work, $4.08 is the top.” Luria’s framework – discount the future agent-labor TAM by an honesty factor – is the new debate on the call sheets of every long-only AI fund.
Historical Context: From $1.10 in 2024 to $4.08 in 2026
To appreciate the scale of the move, rewind two years. In Q1 2024, A100 spot prices on neoclouds traded between $1.10 and $1.80 per hour. By Q3 2024, the H100 was the marquee chip and rented in the $2.20 to $3.50 range on spot – with brief excursions to $4 during Llama-3 training rushes. The H200 launch in late 2024 took spot to $3 and held through most of 2025 as supply caught up. The B200 launch in mid-2024 and rack-scale GB200 NVL72 deployments in late 2025 reset the market entirely.
The pattern is clear: each Nvidia generation has launched at a 1.5-2x rental premium to the previous, and that premium has held longer than analysts forecast. The early 2026 reset broke the pattern in two ways. First, Blackwell did not normalize after launch – it surged in the second half of its first commercial year as agentic inference scaled. Second, the previous-gen H100/H200 did not see the typical 50% post-Blackwell discount; hyperscalers held the line at $7.50, suggesting underlying compute demand has expanded faster than supply on every generation simultaneously. The volatility persisted well past the initial launch window, too – by August 2026, entry-tier RTX PRO 6000 Blackwell spot rentals on platforms like ThunderCompute were still swinging between $1.11 and $4.50 per hour, evidence that pricing never settled into the calm, predictable band prior Nvidia generations eventually found.
The historical comparison most often cited inside Nvidia is the 2017-2019 cryptocurrency mining boom, but the analogy breaks down in two crucial ways: AI compute demand is institutional and contracted, not retail and speculative, and software ecosystems (CUDA, NIM, TensorRT-LLM) lock in customers in a way mining never did. The structural floor under Blackwell rental is therefore higher than under any prior Nvidia cycle.
Expert Voices: Five Reads on the $4.08 Print
Stacy Rasgon, Bernstein: “When a $4 print sticks for two quarters, it’s not a print, it’s a benchmark. We are reading this as a regime change and we expect Nvidia datacenter revenue to print well above consensus for Q2 and Q3.”
J.J. Kardwell, Vultr CEO: “The most severe compute shortage I’ve seen in over five years of running this company. Customer behavior has changed – they are paying any rate to lock in capacity, and they are signing longer.”
Vivek Arya, BofA Securities: “We retain a Buy on Nvidia and now model the GPU supply-demand imbalance persisting through at least 2029. The Ornn print is consistent with our channel checks; we see no near-term mean-reversion catalyst.”
Patrick Moorhead, Moor Insights & Strategy: “AMD has the silicon. What it doesn’t have is the production reference-architecture density that Nvidia shipped with GB200 NVL72. Credibility is not a CUDA kernel –, and that’s why Blackwell pricing remains unhedged for the next 12-18 months.”
Gil Luria, D.A. Davidson: “If Claude Code and Cursor genuinely 5x developer productivity, the capex is cheap. If gains are 20-30% and concentrated in a thin slice of work, $4.08 is the top. The whole bull case is one agentic-disappointment cycle away from rolling over.”
Market Reaction: Nvidia, CoreWeave, AMD, and the Cloud Tier
The April 13 WSJ piece moved markets within hours. Nvidia closed up 3.4% on heavy volume, briefly retaking its prior all-time high; CoreWeave surged 9.1% and printed a fresh 52-week high; AMD added 2.7% on read-through to MI355X/MI400 pricing power, and Vertiv, Eaton, and Schneider Electric – the picks-and-shovels datacenter power and cooling cohort – each closed up between 1.8% and 4.2%. The KraneShares Artificial Intelligence ETF (KOMP) gained 2.1% on the session.
Among the cloud names, Microsoft was flat (the market read its allocation discipline as defensive), Google Alphabet added 1.6% on TPU-substitution optionality, and Amazon ticked up 1.1% on Trainium momentum. Notable underperformers were the application-layer AI software cohort: HubSpot, Salesforce, and ServiceNow lagged the tape by 1-2%, with traders flagging that input-cost pass-through could compress 2026 AI feature margins.
Credit markets noticed too. CoreWeave’s outstanding senior secured notes tightened 18 basis points; the broader datacenter HY index tightened 6 bps. Even the equipment financing market reacted – secondary trading desks reported Nvidia DPU and HGX board collateral commanding higher residual-value assumptions in new lease pricing.
Five Predictions for the Rest of 2026 and Into 2027
1. The $4 floor holds through Q3 2026. CoWoS-L capacity does not meaningfully expand until late Q3, and the new TSMC fabs at Kaohsiung and Arizona Phase 2 will not contribute material throughput until 2027. Until then, Blackwell spot will trade in a $3.80-$5.50 corridor, with brief excursions during model-launch weeks.
2. Three-year commitments become the de facto neocloud unit of trade. Following CoreWeave’s lead, Lambda, Nebius, and Crusoe will move smaller customers to multi-year reservations by Q3, effectively eliminating the consumer-friendly hourly tier for the highest-demand SKUs.
3. AMD MI355X and MI400 take meaningful share – but only at the inference margin. Anthropic, Meta, and a handful of Tier-1 enterprise customers will begin running 5-15% of inference on AMD by year-end, providing the first genuine pricing pressure on Blackwell into 2027. Training stays Nvidia-dominant.
4. Vera Rubin pricing prints north of $6 per GPU-hour on spot at launch. The 336-billion-transistor next-gen platform debuts in late 2026 / early 2027 with 5x Blackwell compute density. The initial allocation will go to OpenAI, Anthropic, xAI, and Microsoft, and the early spot rentals will reset the benchmark again, taking the implicit rental yield on Nvidia silicon higher rather than lower.
5. Sovereign-AI buyers absorb any spare capacity. The UK’s £500M sovereign AI fund, France’s Linux/Mistral push, and the Cohere-Aleph Alpha-Schwarz Group sovereign tie-up will mop up incremental Blackwell allocations through 2026-2027. The result is a structurally tighter merchant market and continued pricing power for neoclouds with sovereign-grade compliance posture.
What This Means for Founders, CIOs, and Investors
For AI-native founders, the playbook has changed. The era of betting on continuous unit-economic improvement through falling inference costs is over for at least the next 18 months. Pricing models that assume token costs decay 30-50% per year need to be re-underwritten; in 2026, blended inference costs are flat to up. The most exposed cohort is the consumer freemium AI startup; the least exposed is the enterprise contract-priced AI vendor that can pass through.
For CIOs and infrastructure leaders, the message is to lock capacity now or accept allocation risk. Three-year Blackwell reservations at hyperscaler list prices look expensive on a sticker basis but compelling on a duration basis if the Ornn floor holds. Multi-cloud Blackwell strategies – splitting between an AWS/Azure reserved baseline and a neocloud burst – are increasingly the modal answer in Fortune 500 architecture reviews. AMD pilots are also moving from optional to mandatory in 2026 procurement budgets.
For investors, the trade is barbelled. The picks-and-shovels names (Nvidia, TSMC, ASML, Vertiv, Eaton, Marvell, Broadcom) are levered to sustained capex. The neoclouds (CoreWeave above all, then Lambda, Nebius, FluidStack) are levered to the spot price and contract-tenor mix. The application-layer AI software stack remains the most vulnerable to input-cost pass-through; the question of whether agentic productivity gains are large enough to absorb $4 rental input costs is the central debate of the second half of 2026.
Frequently Asked Questions
What is the current Nvidia Blackwell GPU rental price in April 2026?
The Ornn Compute Price Index, integrated into Bloomberg Terminal in early April 2026, prints the Blackwell spot rental at $4.08 per GPU-hour as of April 13, 2026 – a 48% jump from $2.75 in mid-February 2026. Hyperscaler reserved Blackwell rates remain meaningfully higher at $10.50-$11.20 per hour on three-year commitments. That volatility only continued further down the product stack: by August 2026, ThunderCompute listed RTX PRO 6000 Blackwell rentals ranging from $1.11 to $4.50 per hour, Spheron offered the same SKU at $2.29 per hour dedicated or $1.19 per hour spot, and Presenc.ai’s Q2 2026 tracker put single B200 rentals as high as $4.50-$7.00 per hour.
Why did Blackwell GPU rental prices surge 48% in two months?
Three forces converged. Agentic-AI inference workloads from Claude Code, OpenAI’s o-series, Cursor, and other agent-heavy products burn 8-40x more output tokens than chat. Post-training and reinforcement-learning sweeps from second-tier model labs absorbed elastic capacity. And on the supply side, TSMC CoWoS-L packaging remains the bottleneck while HBM3e memory prices rose roughly 20% for 2026 delivery contracts.
Which cloud providers are most exposed to the Blackwell price surge?
Neocloud providers – CoreWeave, Lambda, Crusoe, Nebius, FluidStack – are the most direct beneficiaries, with CoreWeave’s roughly 470,000-GPU fleet capturing the largest gross-margin uplift. Hyperscalers (AWS, Azure, GCP) have held public list prices steady but are tightening allocation to Enterprise Agreement customers, pushing smaller buyers onto longer reservations.
Are AMD and Google chips cheaper alternatives to Nvidia Blackwell?
Yes on a per-hour basis – AMD MI300X spot trades around $2.10-$2.50 and Google TPU v6e prints in the $2.70-$3.10 band – but software-ecosystem gravity (CUDA, NIM, TensorRT-LLM) and multi-year Nvidia purchase commitments mean substitution remains slow. AMD MI355X and MI400, plus TPU v7p, are expected to capture meaningful inference share by year-end 2026, but training stays Nvidia-dominant.
How long will the Blackwell shortage last?
Bank of America’s semiconductor team projects the GPU supply-demand imbalance persists through at least 2029. TSMC’s $56 billion 2026 capex aims to roughly double CoWoS-L capacity to a 75,000-wafer-equivalent monthly run rate by year-end, but new lines do not light up until late Q3 2026. Expect Blackwell spot to trade in a $3.80-$5.50 corridor through Q3.
What does the surge mean for OpenAI, Anthropic, and other AI labs?
Real pain. OpenAI shuttered consumer Sora to free Blackwell capacity for the API. Anthropic missed its 99.9% Claude API uptime SLA across Q1, triggering enterprise SLA credits. Cursor, Replit, Factory AI, and other inference-heavy startups raised effective per-seat prices 30-40% on a blended basis. Meta’s $135 billion 2026 capex is buying fewer effective hours than at the December 2025 commit.
Is this an AI capex bubble?
The bull case treats $4.08 as validation of the $650 billion big-tech capex cycle. The bear case worries that agentic productivity gains are not yet large enough to absorb the input-cost pass-through, and that a disappointing agent demand cycle could collapse rental rates and leave neoclouds with overpriced depreciating silicon. The debate will likely be resolved by mid-2027 earnings.
How does $4.08 compare to historical GPU rental prices?
In Q1 2024, A100 spot prices were $1.10-$1.80 per hour. H100 spot peaked around $4 in 2024 training rushes and settled to $2.20-$3.50 through 2025. Blackwell’s $4.08 spot print is a roughly 2x premium to H100 at the same point in its life cycle, and unlike prior cycles the previous-gen H100 has not seen a typical post-launch discount – suggesting structurally tighter compute demand across every Nvidia generation.
Related Coverage
- CoreWeave’s Anthropic Deal: 12% Surge, $6.8B Backlog [2026]
- Nvidia Vera Rubin Platform: Inside the 336B-Transistor Chip and 5x Blackwell Leap
- AMD’s MI400 Series: Inside the 320B-Transistor Chip and $7.2B Bet to Break Nvidia’s Grip
- Google’s TPU 8t and 8i: Inside the 121 Exaflops Bet and $21B Broadcom Deal
- Big Tech’s $650 Billion AI Infrastructure Bet
- TSMC’s $35.71B Q1 2026 Revenue: Inside the 35% Surge and $56B Capex
- AI Chips 2026: The Complete Buyer’s Guide
External Sources
- Silicon Data: B200 Rental Index – March 2026 Update
- CoreWeave Corporate Site
- Nvidia HGX Platform Specifications
- Tom’s Hardware GPU Coverage
- Anthropic Developer Updates
Published April 13, 2026. Data current as of the Ornn Compute Price Index reading dated April 13, 2026 and the Silicon Data SDB200RT index dated March 30, 2026.
