Is the AI Chip Shortage Over in 2026? GPU Pricing and HBM Backlogs

H100 cloud rental rates have fallen from roughly $8/hr in 2023 to $1.80-3.50/hr in Q2 2026 — but reserved contract pricing has climbed nearly 40% since October 2025. That's the short answer. The longer answer is that the AI chip shortage didn't end, it moved.

Every few months someone declares the GPU shortage "over" because spot prices dipped. Then reserved pricing climbs back up two quarters later and everyone acts surprised. Having watched this cycle from the venture side since 2023, the pattern is now clear: raw GPU die supply genuinely improved through 2025, but the constraint just shifted one layer down the stack, from the processor itself to the HBM memory stacked on top of it. That distinction matters enormously if you're a founder budgeting compute costs or an LP trying to underwrite the next infrastructure fund.

Is the AI Chip Shortage Over in 2026?

Not entirely. On-demand H100 rental rates have dropped roughly 60-77% from their 2023 peak of about $8/hr to $1.80-3.50/hr in Q2 2026, with spot pricing as low as $1.20/hr at the cheapest providers. But reserved 1-year H100 contract pricing bottomed at $1.70/hr in October 2025 and has since climbed nearly 40% to $2.35/hr by March 2026, and lead times for new data-center GPU capacity still run 36-52 weeks.

How Much Have GPU Rental Prices Fallen Since 2023?

The headline decline is real: H100 hourly rental rates have fallen from roughly $8/hr in early 2023, when the ChatGPT-driven compute rush began, down to a $1.80-3.50/hr range across most providers in Q2 2026. That's the number that generates "shortage is over" headlines. But it obscures two things — reserved contract pricing (what most serious AI labs and startups actually pay) bottomed out in October 2025 and has been climbing since, and the newer Blackwell generation is pricing at a significant premium to Hopper.

Why Are H100 and H200 Prices Climbing Back Up in 2026?

H100 and H200 contract pricing climbed roughly 40% between October 2025 and March 2026, and the driver isn't GPU die scarcity — it's HBM3e memory cost pass-through from Samsung and SK Hynix. HBM3e prices for 2026 delivery contracts rose about 20% year-over-year, and since every H200 and Blackwell package requires multiple HBM stacks bonded via TSMC's CoWoS packaging process, that memory cost flows straight into what cloud providers charge for reserved capacity. Roughly half of the providers tracked by industry surveys report zero available Hopper-generation capacity coming off contract at all.

What's Actually Causing the 2026 AI Chip Shortage — GPUs or HBM Memory?

HBM memory, not the GPU compute die itself, is the binding constraint on AI chip supply in 2026. TSMC's CoWoS-L advanced packaging capacity — required to bond HBM stacks onto every H200 and Blackwell substrate — is fully allocated through at least mid-2027. Samsung and Micron are both ramping HBM3e output, but neither is expected to meaningfully ease the shortage before late 2026 at the earliest, which means the "chip shortage" framing is now more accurately a memory-and-packaging shortage riding on top of an otherwise recovering GPU die supply chain.

How Much Does It Cost to Rent an Nvidia Blackwell B200 GPU in 2026?

Nvidia B200 rental rates run roughly $4.50-7.00/hr at specialist neoclouds in Q2 2026, having jumped to $4.08/hr in April alone — a 48% increase in just two months as Blackwell demand outpaced allocation. Hyperscaler list pricing runs much higher: AWS's Blackwell instance rate reaches about $14.24 per GPU-hour on-demand, though three-year reserved commitments bring the effective hyperscaler rate down to roughly $10.50-11.20/hr. GB200 NVL72 rack-scale systems remain tightly allocated and are mostly accessible only through hyperscaler clouds or a handful of top-tier neoclouds.

How Long Are GPU Lead Times in 2026?

Data-center GPU lead times still run 36-52 weeks as of mid-2026, only a modest improvement from the worst stretches of the 2023-2024 shortage. Roughly half of cloud providers surveyed by industry trackers report being completely sold out of Hopper-generation capacity, with no capacity expected to free up from expiring contracts anytime soon. For a startup planning a training run 12 months out, that lead time is still long enough to require locking in reserved capacity well before the compute is actually needed — which is exactly the dynamic pushing reserved pricing higher even as spot rates fall.

What Does Nvidia's Earnings Data Say About AI Chip Demand Going Into 2027?

Nvidia's fiscal Q2 2026 revenue rose 56% year-over-year to $46.74 billion, and data center revenue grew at the same 56% clip despite a $4 billion hit from China export restrictions on H20 chips. Nvidia has told investors it's tracking roughly $1 trillion in confirmed AI chip purchase orders through 2027 — not analyst projections, but signed demand — and guided Q3 FY2026 revenue to about $54 billion, plus at least 42% forward revenue growth over the next year. That's not the earnings profile of a company facing a demand cliff; it's a company still supply-constrained on the memory and packaging side of its own product.

Q2 FY2025 prior-year figure is a directional estimate derived from Nvidia's reported 56% YoY growth rate; Nvidia does not restate prior-year figures in the same release.

GPU Pricing Across Providers and Chip Generations, 2026

Figures are Q2 2026 estimates blended from SemiAnalysis' H100 Rental Price Index, IntuitionLabs' Data Center GPU Pricing 2026, Thunder Compute's GPU Rental Market Trends (July 2026), and Spheron's CoreWeave pricing analysis. Rates vary by commitment length, region, and bundle size.

None of this happens in isolation from the broader capex story — hyperscalers are still pouring hundreds of billions into data centers precisely because compute remains the gating factor on model training and inference at scale, a dynamic we track on the Big Tech Earnings Dashboard and the AI Valuations Tracker.

The Bottom Line

If you're budgeting AI compute for 2027, don't anchor on the falling spot-price headlines — anchor on the 36-52 week lead times and the 40% run-up in reserved pricing, because that's what you'll actually be quoted if you need guaranteed capacity. The chip shortage of 2023-2024 was a GPU die problem that the industry mostly solved. The shortage running through 2026 is a memory and packaging problem, and TSMC's CoWoS allocation through mid-2027 means it isn't clearing on the same timeline.