Alongside the launch of our H100 1-Click Rental Index, we wrote up what the GPU rental market actually looks like in early 2026, and the headline is that the spot market for compute has gone from "finally cooling off" in October to a hard squeeze again, in roughly five months.… pic.twitter.com/XayxYXWcZ6
— SemiAnalysis (@SemiAnalysis_) June 12, 2026
Current Live Marketplace Prices (Cheapest Public Options)
Vast.ai (peer-to-peer marketplace — mix of interruptible/spot and on-demand)
B200 (192GB HBM3e)
Cheapest listings: ~$3.44/hr
Typical range: $3.44 – $7.45/hr
Median/typical: ~$4.5–4.7/hr (p25 ~$4.23, p90 ~$7/hr)
High availability.
B300 (Blackwell Ultra, 288GB)
Cheapest listings: ~$5.33/hr
Typical range: $5.33 – $7.33/hr (up to ~$8.27 at p90)
Median/typical: ~$6.2/hr
Medium availability.
RunPod (managed cloud, more reliable SLAs)
B200 (180–192GB): ~$5.89/hr (community/on-demand Pods).
B300 / HGX B300 (288GB) ~$6.94/hr.
Broader Market Context (Across Providers)
B200 medians Often $5–6/hr (ranges from ~$3.75 on cheaper neoclouds to $14+ on hyperscalers like AWS).
B300 medians Often $6–7/hr (ranges from ~$6.10 to $18 on hyperscalers like Oracle).
Hyperscalers (AWS, Google Cloud, Oracle, etc.) Significantly higher — frequently $10–16+/hr for on-demand/spot B200.
Longer-term reserved contracts can be lower than pure on-demand but still elevated due to tightness.
Spot/interruptible (cheaper, can be preempted) is usually the lowest on Vast.ai.
On-demand/guaranteed costs more for uptime.
Availability is tighter for B300 (higher-end Blackwell Ultra) than B200.
Spot prices are generally 40–70% cheaper than on-demand rates but come with preemption risk.
Reserved contracts offer meaningful discounts versus on-demand (often 20–50%) with guaranteed availability and higher priority.
Brian Wang is a Futurist Thought Leader and a popular Science blogger with 1 million readers per month. His blog Nextbigfuture.com is ranked #1 Science News Blog. It covers many disruptive technology and trends including Space, Robotics, Artificial Intelligence, Medicine, Anti-aging Biotechnology, and Nanotechnology.
Known for identifying cutting edge technologies, he is currently a Co-Founder of a startup and fundraiser for high potential early-stage companies. He is the Head of Research for Allocations for deep technology investments and an Angel Investor at Space Angels.
A frequent speaker at corporations, he has been a TEDx speaker, a Singularity University speaker and guest at numerous interviews for radio and podcasts. He is open to public speaking and advising engagements.
How many AI data centers are coming online by the end of 2027?
This is from Stargate in Abilene, Texas:
Current (as of ~April/May 2026 reports): Roughly 4 of the planned 8 buildings are operational, with ~0.3 GW total facility power and compute equivalent to roughly 250,000 H100s (Blackwell GPUs are significantly more powerful per chip than H100s).
Full Abilene campus (targeted for completion around mid-to-late 2026): 8 interconnected buildings, 1.2 GW total power, and hundreds of thousands of GPUs — with reports citing plans for 450,000+ NVIDIA GB200 GPUs under Oracle’s lease.
This is just one of many expected AI datacenters to come online.
Goldman Sachs (using granular facility-level data) forecasts massive US capacity additions: ~13.6 GW in 2026 + 36.3 GW in 2027, pushing total US data center capacity toward ~95 GW by end-2027. This implies hundreds of individual facilities or campus phases activating over those two years.
Just thought some readers might like to discuss this in context.
Anthropic/XAi rental rate is ~$7.78/GPU‑hr vs H100 on-demand rates around $3 to $4 per GPU per hour. (new-clouds offer rates down to $1.49–2.99. )
Anthropic is paying a temporary overprice as a stop gap measure. This will end as soon as the new breed of inference hardware comes online. H100 is basically obsolete already so the current deal is very good for XAi. It won’t last long though.