buildConfirmed2 publishers Ai2 released Olmo-core 3, an open mixture-of-experts training stack it has benchmarked at over one trillion total parameters. Its speedups are measured against Ai2's own earlier FSDP code, so teams on Megatron-Core must run their own comparison.
Perspective Coverage
3 publishers
- Builder
- Builder 55%
- Operator
- Operator 25%
- Investor
- Investor 20%
Reality
- Evidence50
- Adoption
- Not scored
- Hype gap+20
- Incentives55
- Confidence60
buildConfirmed3 publishers Semi-Tech Leasing Group, a state-backed Chinese lessor, financed 32 Asus servers with restricted Nvidia B300 chips for Hongjing Technology, Bloomberg reports. Hongjing's filings confirm the Semi-Tech leases, so a lender sits on the path from manufacturer to operator.
Perspective Coverage
4 publishers
- Builder
- Builder 11%
- Operator
- Operator 39%
- Investor
- Investor 50%
Reality
- Evidence62
- Adoption
- Not scored
- Hype gap+20
- Incentives60
- Confidence58
buildOne report1 publisher Modal's gang scheduler places multi-node GPU training jobs from a shared serverless pool
Modal made multi-node GPU clusters generally available on October 1st, requested through one Python decorator and billed by the second. Dropping a reserved cluster for it means trusting a gang scheduler to place every node at once on one network.
Reality
- Evidence45
- Adoption35
- Hype gap+20
- Incentives65
- Confidence50
buildOne report1 publisher FastGPU finds hyperscalers charging three times the market floor for H100 and B200 hours
FastGPU's September 27 snapshot of 28 GPU clouds puts the cheapest hyperscaler H100 at $5.38 an hour, 3.0x the $1.79 market floor. That premium buys IAM, managed services and credits, and it is worth paying when a team actually uses them.
Reality
- Evidence45
- Adoption
- Not scored
- Hype gap+15
- Incentives65
- Confidence50
buildConfirmed2 publishers Moonshot published 2.8 trillion open weights. At four bits per parameter that is about 1.4TB resident before any cache, which rules out the eight-way H100 node most teams assume.
Reality
- Evidence60
- Adoption
- Not scored
- Hype gap+10
- Incentives40
- Confidence58
The Atlas 960E fits 8 EFLOPS behind Huawei's own optical engines, and the Ascend 960DT arrives three quarters early, while rotating chairman Eric Xu says supply goes to Chinese customers first and dates China's catch-up with its own demand to 2030.
Perspective Coverage
4 publishers
- Builder
- Builder 26%
- Operator
- Operator 38%
- Investor
- Investor 36%
Reality
- Evidence55
- Adoption50
- Hype gap+30
- Incentives70
- Confidence60
buildOne report1 publisher White Circle publishes the harness behind Halo's 2.3x TRL throughput claim
Halo adds expert and tensor parallelism to Hugging Face models and still saves SafeTensors that from_pretrained can load. Its best number, 9,009 tokens per second per GPU against TRL's 3,885, came from synthetic fixed-length sequences.
Reality
- Evidence45
- Adoption14
- Hype gap+22
- Incentives72
- Confidence56
buildOne report1 publisher NVFP4 squeezes Qwen3.8's 2.4 trillion weights onto eight B300s at 150 GB a GPU
Alibaba's largest open-weight release fits on a single eight-GPU node only because a community four-bit build takes it down to 1.2 TB. AWS publishes the vLLM config for it and skips the price.
Reality
- Evidence52
- Adoption22
- Hype gap+32
- Incentives78
- Confidence48
buildOne report1 publisher NVIDIA's 1.653-second H3 clip is built on a four-step adapter, with faster kernels layered on top
NVIDIA Research rebuilt the serving stack for someone else's open-weight video model and got a five-second clip out in 1.653 seconds, but the default profile is lossy and the dense comparison run is the honest baseline.
Reality
- Evidence62
- Adoption45
- Hype gap+30
- Incentives72
- Confidence58
buildOne report1 publisher Eight B300s finish a 10.125-second MiniMax clip with 1.4 seconds of headroom
The vLLM benchmark shows a finished MP4 arriving before its own playback would end, which is a different property from showing frames as they are made, and MiniMax's community licence still requires separate permission for US, EU, UK and Korean use.
Reality
- Evidence58
- Adoption38
- Hype gap+38
- Incentives74
- Confidence55
Taiwan's indictment over 130 Nvidia B300 servers rests on forged paperwork and a staged site visit. The defendants are salaried vendor staff, not brokers, and the charges are personal.
Perspective Coverage
3 publishers
- Builder
- Builder 12%
- Operator
- Operator 50%
- Investor
- Investor 38%
Reality
- Evidence74
- Adoption58
- Hype gap+12
- Incentives66
- Confidence71
Together AI rents India's largest GPU cluster rather than building it
L&T's Vyoma unit will host 10,000 Nvidia B300s at a 250MW first phase in Chennai. The tenant is a US AI cloud whose own sites are in North America.
Reality
- Evidence42
- Adoption24
- Hype gap+38
- Incentives78
- Confidence55