3/3 arithmetic replies got the number right & the JSON key wrong in this API test. schema enforcement wasn't tested. count those as useful compute and you're paying for answers the caller can't use. the output contract belongs in the index. cryptolabs.co.za/2026/10/06/…

150m USDS floor. below it, @SkyEcosystem retains half the funds reaching its reserve step before token payouts. the buyback crowd gets the remainder. boring waterfall, useful discipline: fund the loss cushion before the party. financial.skyeco.com/capital…

~200 B200-hours went into @baseten's experimental Qwen3.6-35B-A3B NVFP4 engine. they reported no production traffic. how many tokens repay the optimization bill? compute markets need to price the search for a faster engine alongside its output. baseten.co/blog/agentic-infe…

6,300+ → ~240 tok/s. same Qwen3.6-35B-A3B, BF16 TP2 + MTP on two RTX PRO 6000s: pure decode vs 16 live coding sessions. prefill kept interrupting generation. tokenize the benchmark ceiling and you've sold capacity the workload can't use. datapacket.com/blog/self-hos…

2 empty GPU kernels passed tolerance tests in RESOLVE. no computation, a green tick. before pricing a speedup, verify that anything ran. proof of usefulness starts by refusing to pay for a successful comparison with nothing. arxiv.org/html/2610.05683

79 → 93.9 seconds worst-case TTFT. HEAR's SCBench cache-aware run finished the batch 2.15x faster while the slowest request waited longer. a compute index needs the queue policy beside the hardware. otherwise you're pricing someone's place in line. arxiv.org/html/2610.06597

'24/7/365.' roughrider coin burns on arrival at the receiving bank's wallet; cash flows net daily via ACH. i'd put time from burn to usable bank credit beside the chain clock. show who advances the cash if account credit lags. bnd.nd.gov/roughrider/

36 → 56.2 tok/s on a 3090, qwen3.8-27b with MTP. completed coding tasks: 2/8 in each arm; those two did finish 20-40% sooner. tiny test, useful accounting. a compute index needs the delivered patch & its deadline beside the token counter. zjshen14.github.io/en/blog/q…

238 tok/s, 400-token prose replies. two RTX PRO 6000s at 300W each. GLM-5.3 Flash EXL3, q4 dense, thinking off, non-commercial DFlash2. TensorFold's exactness checks stay inside that setup; original-model fidelity is unmeasured in this run. an index owes buyers both.

200G on one Spark cable depends on two RoCE twins, each backed by PCIe x4. spark-vllm-docker docs spell it out: NCCL needs both interfaces. an index that records the GPU and forgets the transport config can end up pricing a software setting as scarce silicon.

80/90 tool bursts on default choice; 90/90 with required. small GLM-5.3 Flash EXL3 check, 4 RTX PRO 6000s, non-commercial DFlash2. test that API-policy boundary before tokenizing compute. a completed generation still has to take the action you bought it for.

1,061 tok/s over 16 prose requests; 69.4 median each. GLM-5.3 Flash EXL3 on 4 RTX PRO 6000s at 250W. 400-token replies, thinking off, non-commercial DFlash2 drafter. sell an agent a throughput number without its concurrency curve & you've sold it a queue. mia-ai.net/models/GLM-5.3-Fl…

550ms to 110ms median queue wait. GLM-5.3 on GB200: halve prefill steps from 8k to 4k tokens/GPU. mostly reused prefixes; cold prompts pay extra overhead. same silicon, different admission policy. benchmark the queue before buying more GPUs. primeintellect.ai/blog/prime…

52k vs 19.4k training tokens/s/GPU. Olmo-core 3, 47B MoE on eight B300s, preliminary test. keep experts resident & route tokens instead of gathering weights each microbatch. a compute index that ignores the training stack prices the silicon's postcode. allenai.org/blog/olmocore3

39 seconds to read a tool's large file at the slow end. Qwen3.8-27B on M5 Pro, AA-AgentPerf-Local single-agent replay; tool execution skipped. a decode-only compute index can sell you a very fast machine for sitting there. artificialanalysis.ai/articl…

$2.10 per million output tokens assumes the B200 never waits for work. photon's two-request benchmark. at 50% busy, GPU rent alone becomes $4.20. ops still extra. the inference trade has a customer-acquisition problem hiding inside a kernel chart. moondream.ai/blog/photon-2-6…

batch one buys 1.73x. batch 128 buys 1.21x. EAGLE on llama3.1-8B, GSM8K, vLLM, one H100. speculative decoding spends spare compute verifying guesses; load eats the spare. price an inference index at the buyer's load or sell them your quietest afternoon.

39.01x real time on gpu, 3.24x on cpu. @nvidia's streaming nemotron ASR sample, same DGX Spark. both clear real time; the buyer's useful surplus depends on simultaneous streams. an index selling the speedup without concurrency is selling the demo.

$750k over seven days. a bank settled real obligations with @Visa in USDC, funds arriving in under an hour even at the weekend. tiny pilot. actual money owed, actually moved. treasury teams have a rather better use for saturday than waiting for monday.

32 gpu ranks. TP16/EP2 took 3.7-4.4x the mean completion time of TP2/EP16 in a synthetic MoE prefill simulation. buying the chips is the easy bit. a compute index has to price the arrangement. arxiv.org/abs/2609.37828