Luxor's Compute NewsletterWEEKLY
Weekly coverage of the AI compute market: the news, the numbers, and the infrastructure economics that matter.
August 26, 2026 · 5 min read
Get every issue in your inbox
No spam. Unsubscribe anytime.
Compute Price Pulse
1-yr contract · median $/GPU-hr
| H100 | $2.97 ▲ 8.4% since last week |
Available on Hashrate Index. Powered by Ornn.
The Lead
A GPU running AI inference just answered a live grid signal
Earlier this month, a GPU serving AI inference took a Four Coincident Peak curtailment signal from Luxor's Qualified Scheduling Entity and cut its power draw to roughly a quarter of normal levels in under 500 milliseconds. Bentaus' Ziani software ran the whole loop, signal to verified curtailment, with no operator in the middle. The inference jobs kept running.
Texas miners have played 4CP for years, because an ASIC is the easy case: power it off and you lose uptime and nothing else. A GPU stopped mid-job can lose the work in progress, which is why AI data centers get modeled as inflexible baseload in interconnection queues. Checkpointing closes that gap, and this is the first evidence it holds up against a real grid signal instead of a whiteboard. Luxor and Bentaus are now standing up a cluster node to chase additional ERCOT programs and settle the economics: does a GPU cluster earning program revenue beat one running flat out at full uptime?
The SPOTLIGHT
Bitdeer AI sold half a site before it turned on
Bitdeer AI has deployed GB300 NVL72 capacity at a Malaysian facility it calls A102. The site will offer 9.5MW once operational, and half of it is already committed under a five-year offtake the company values at $400 million, with revenue and cost effects landing in Q1 2027. Negotiations for the rest of the capacity are ongoing. Bitdeer AI describes A102 as a multi-customer building purpose-built for rack-scale liquid cooling, and is targeting 350MW of AI cloud capacity by Q1 2028. The parent company still has 175,000 Bitcoin miners under management and is converting former mining sites in Washington, Tennessee, and Norway to AI.
LUXOR'S TAKE
The megawatts are not the story. The contract is. Selling half a building before energization is what separates an AI pivot from an AI business, and it is the bar every operator converting mining capacity should expect to clear before the racks arrive.
AI Hardware availability
B300 and H200 in stock — start the revenue clock sooner
When the neocloud math works, sourcing is the next question. Luxor has B300 and H200 GPU hardware available now.
Around the Industry
Throughline: compute is starting to get priced, hedged, and squeezed like a commodity instead of bought like equipment.
The CFTC is asking how to regulate compute derivatives
The commission opened a public comment request on compute derivatives contracts, seeking input on compute cash markets, market oversight, manipulation and customer protection concerns, and perpetual compute futures. Chairman Michael Selig called it a first step toward clear rules for American compute markets. (Reuters)
Keel has switched off every one of its US mining sites
For site operators
The company formerly known as Bitfarms shut Moses Lake in April and its three Pennsylvania sites in June, and is now pushing three US data center developments worth a combined 478MW of gross capacity. Moses Lake should come online first in 2027. Two of the retired sites still earn money selling electricity. (Data Center Dynamics)
A startup says deployed GPUs have capacity left in them
Vectris Labs claims it can recover productive capacity from GPUs that are already racked, and will launch a control layer called Waveform to design partners on October 1. Its own tests on rented H100, H200, and B200 instances report throughput gains between 30% and 73% with lower energy use per job, no retraining or kernel changes required. (Vectris via PR Newswire)
Chip suppliers keep buying into the companies that buy their chips
NVIDIA is reportedly in talks to take an equity position in Perplexity at a valuation above $30 billion, after first exploring a licensing arrangement. Neither company has confirmed the discussions. The pattern matters more than the number: when the supplier of the silicon also owns a piece of the applications consuming it, demand for compute stops being something you forecast and starts being something you underwrite. (Forkast)
The AI data center is not going to run on one kind of chip
A roundtable of chip and design engineers makes the case that clusters are moving toward a mix of CPUs, GPUs, NPUs, and custom accelerators, with the inference pipeline split so different stages run on the hardware that suits them. Power is the binding constraint and cost per token is the scoreboard, which is what pushes operators toward hardware diversity in the first place. The hard part becomes the orchestration software once the silicon arrives from several vendors at once. (Semiconductor Engineering)
From the Vault
Evaluating Bitcoin mining sites for AI training & AI inference
Training and inference are two different infrastructure markets, and your site almost certainly fits one better than the other. Five gating questions covering location, power, cooling, networking, and operations, plus an interactive evaluator that scores the site before you start procurement conversations. Read now →
AI Hardware availability
Offtake is becoming a financed product.
Luxor’s compute offtake listings show real-time availability — fill capacity or secure it.
