Appendix
Raw scores by run
These raw scores read as direction, not level - the words on the main page are the honest summary.
Every ranked signal this run
Sharon AI signed a six-year computing deal with Nvidia to build 72 megawatts of new data-center capacity in Australia and deploy up to 40,000 of Nvidia's Grace Blackwell GB300 chips, with more than 55,000 Nvidia GPUs total expected to be running by mid-2027.
observed 2026-06-12 - SharonAI Holdings (SEC Form 8-K exhibit 99.1) (primary)
Meituan built its LongCat-2.0 AI model — which has 1.6 trillion internal settings and only switches on part of itself at a time, a design called mixture-of-experts — entirely on more than 50,000 Chinese-made custom AI chips (not Nvidia's), linked together in large computing clusters, training on more than 35 trillion chunks of text. This shows that a top-tier AI model can be trained on hardware other than Nvidia's.
observed 2026-06-30 - Meituan LongCat (secondary)
The Decoder separately confirms that Meituan trained its LongCat-2.0 model (1.6 trillion internal settings) on more than 50,000 Chinese-made custom AI chips with no Nvidia chips involved. Meituan also claims the model does well on coding tests, though the model itself is not widely available to the public yet.
observed 2026-06-30 - The Decoder (secondary)
Taiwanese authorities stepped up their investigation into Nvidia GPU smuggling, raiding Supermicro's Taiwan office and related sites and detaining employees. This is tied to US charges alleging that about $2.5 billion worth of servers were shipped to China without the required export license.
observed 2026-07-01 - Data Center Dynamics (secondary)
Nvidia is expanding a program where it financially backs customers' GPU purchases in exchange for a share of their revenue, and agrees to rent back any unused GPUs at fixed rates. Early customers in the program are Firmus (170,000 GPUs) and Sharon AI (40,000 GB300 chips) — about 210,000 GPUs combined — following earlier support deals with CoreWeave ($6.3 billion) and Lambda ($1.5 billion).
observed 2026-07-02 - Data Center Dynamics (secondary)
The Nvidia-Sharon AI deal is set up so Nvidia earns its normal product revenue plus a cut of Sharon AI's cloud revenue, which helps a cash-constrained customer afford Nvidia's hardware. Unlike some other Nvidia deals, the filing includes no promise for Nvidia to rent back unused GPUs, financially back the purchase, or buy the equipment back later.
observed 2026-06-12 - SharonAI Holdings (SEC Form 8-K exhibit 99.1) (primary)
Memory makers are reportedly shifting production away from standard memory chips toward high-bandwidth memory for AI chips, which is tightening overall memory supply and driving up the memory costs inside Nvidia's GPU and AI-server parts lists.
observed 2026-06-29 - AppleInsider (secondary)
The cost of parts for Nvidia AI-server racks is rising fast — about 2% to 3% per week, already 10% to 15% above normal, and sometimes jumping as much as 40% in a single week — because memory chips and fast storage drives are in short supply.
observed 2026-07-04 - BigGo Finance (secondary)
Nvidia's financial-backing and revenue-sharing deals may mean more and more of its GPU demand comes from GPU rental clouds that Nvidia itself is helping to fund — raising the risk of "circular financing," where Nvidia's own money is what's generating the sales it reports as demand.
observed 2026-07-02 - Data Center Dynamics (secondary)
A lawsuit filed in California alleges that Samsung, SK hynix and Micron — which together control nearly 95% of the standard memory chip market — are holding back production of standard memory chips while shifting their factories to make higher-profit high-bandwidth memory for AI servers instead, keeping memory supply tight for AI and GPU builds.
observed 2026-06-30 - Gizmochina (secondary)
Lambda's GPU rental cloud charges $6.69 per hour per chip for pay-as-you-go access to Nvidia's B200 GPU, in systems with 8 of them installed together.
observed 2026-07-06 - Lambda (secondary)
Lambda's GPU rental cloud charges $3.99 per hour per chip for pay-as-you-go access to Nvidia's H100 GPU, in systems with 8 of them installed together.
observed 2026-07-06 - Lambda (secondary)
Lambda's GPU rental cloud charges $2.79 per hour per chip for pay-as-you-go access to Nvidia's 80GB A100 GPU, in systems with 8 of them installed together.
observed 2026-07-06 - Lambda (secondary)
CoreWeave charges $6.50 per hour for pay-as-you-go access to a single Nvidia GH200 chip in North America.
observed 2026-07-06 - CoreWeave (secondary)
Nebius raised its pay-as-you-go GPU cloud rental prices by about 30% on June 1, reflecting a shortage of AI computing power.
observed 2026-07-04 - BigGo Finance (secondary)
Run history
- 2026-07-02: 17 findings, 8 sources
- 2026-07-03: 20 findings, 6 sources
- 2026-07-05: 10 findings, 9 sources
- 2026-07-06: 15 findings, 9 sources
Back to the page