Recent
- News briefCP-2026-0144
A bitcoin miner leases its whole campus to an AI lab
Riot has leased all 910 MW of its Corsicana campus to an unnamed frontier AI lab for 15 years, about $28 billion in contracted rent.
2 minSource: SEC EDGAR
- News briefCP-2026-0143
A $115 billion bet that AI's wiring turns to light
Marvell will buy laser maker Lumentum in an all-stock deal valuing it at $115 billion, betting that AI connectivity goes optical.
2 minSource: Marvell
- News briefCP-2026-0142
Apollo buys the GPUs, SpaceX pays over five years
An 8-K describes a finance lease: an Apollo-managed entity buys GPU pods, and SpaceX repays about $40 billion of principal quarterly.
2 minSource: SEC EDGAR
- AnalysisCP-2026-0141
A hunch said 25 gigawatts; the concrete says more
SemiAnalysis answers a public doubt about its 2027 buildout forecast by pointing at sites already past halfway.
2 minSource: SemiAnalysis
- AnalysisCP-2026-0140
Patent filings describe a bonder on a track
SemiAnalysis reads ASML's filings, roadmap and a news report as a die-to-wafer hybrid bonding tool in progress.
2 minSource: SemiAnalysis
- News briefCP-2026-0139
An FTX-era bitcoin miner is being turned into AI capacity
SWI Group raised €800 million at a €6.1 billion valuation and plans to move Genesis Digital's US and Swedish sites to AI work.
2 minSource: CoinDesk
- Market briefCP-2026-0138
China's datacentre fleet, counted by building
SemiAnalysis tracks more than 24GW across 1,000+ Chinese facilities, above EMEA, with ByteDance in roughly a fifth of delivered capacity.
2 minSource: SemiAnalysis
- AnalysisCP-2026-0137
What has to stay close together
Mixture of experts did not only raise parameter counts. It changed which tensors are active per token, and therefore which links in a cluster have to be fast.
3 minSource: SemiAnalysis
- Field reportCP-2026-0136
Grass: 74 days on a bandwidth node, and what it produced
Grass turns the internet connection you already pay for into a share of the AI data market. We ran a node, measured it, and recommend it.
6 min
- AnalysisCP-2026-0135
Seven hundred watts per compute die
SemiAnalysis puts numbers on OpenAI's Jalapeno: 13.4 PFLOPs of MXFP4, 15.4TB/s of HBM4, and a volume ramp that only arrives late in 2027.
3 minSource: SemiAnalysis
- AnalysisCP-2026-0134
Same wafer, twice the clock, three per rack
Cerebras doubles CS-3 throughput without a new chip — and takes rack power from 23kW to about 130kW to do it.
2 minSource: SemiAnalysis
- BenchmarkCP-2026-0133
Same eight GPUs, thirty-three points more
Replacing a first-in-first-out scheduler with a constraint-aware allocator moved utilisation from 53.6% to 87.0% on identical hardware.
2 minSource: Hugging Face
- AnalysisCP-2026-0132
A modelling error that cost $11.57 billion
PJM undercounts the plants it already has by about 4GW, and the emergency auction it has scheduled repeats the assumption.
3 minSource: SemiAnalysis
- Field reportCP-2026-0131
Decentralised GPU supply finally met real demand — unevenly
Two years of idle capacity ended this quarter for some networks and not for others. The split says more than the headline.
12 min
- Market briefCP-2026-0126
Spot pricing is not the differentiator anyone claims
Headline hourly rates converged. Availability at the advertised rate did not.
5 min
- AnalysisCP-2026-0120
Verification is the hard part, not scheduling
Proving a job ran as specified remains unsolved at acceptable cost.
9 min
- AnalysisCP-2026-0114
Storage networks are quietly the profitable half
Compute gets the attention. Decentralised storage got the paying customers first.
6 min
- Research noteCP-2026-0108
Power, not silicon, is the binding constraint
Accelerator supply loosened. Grid connections did not.
7 min