The 25 million shares vest only against purchases Amazon actually makes, so Amazon's downside is nil and the $60bn is a ceiling with no published floor beneath it. What has to show up is the $5bn data-centre guide for fiscal 2027.
Perspective Coverage
7 publishers
- Builder
- Builder 24%
- Operator
- Operator 20%
- Investor
- Investor 56%
Reality
- Evidence70
- Adoption30
- Hype gap+35
- Incentives60
- Confidence65
buildOne report1 publisher FBTriton says its Triton embedding-lookup kernels outperform legacy CUDA
FBTriton's authors say their Triton rewrite of Table Batched Embedding beats legacy CUDA kernels, and an optional B200 path cuts combined latency 16.8%. That path is off by default, and CUDA still validates inputs before the Triton forward pass runs.
Reality
- Evidence40
- Adoption
- Not scored
- Hype gap+30
- Incentives65
- Confidence45
buildConfirmed6 publishers Aleph Alpha released Kolibri, an Apache 2.0 German-English model that activates 3.46B of its 78.1B parameters per token. Each token costs about as much compute as a small model, yet a team hosting it in Europe still has to fit every expert in memory.
Perspective Coverage
6 publishers
- Builder
- Builder 47%
- Operator
- Operator 37%
- Investor
- Investor 16%
Reality
- Evidence72
- Adoption
- Not scored
- Hype gap+15
- Incentives65
- Confidence70
CME's 36-month GPU futures put a forward price on the compute bitcoin miners now sell
CME plans to list two Nvidia GPU compute futures on October 5, pending regulatory approval, with prices running 36 months forward. Bitcoin miners that moved into AI hosting can hedge with them only if what they are paid moves with a GPU rental index.
Reality
- Evidence42
- Adoption10
- Hype gap+55
- Incentives78
- Confidence45
buildOne report1 publisher Meta's 3.2K-line Triton kernel outruns FlashAttention-4 on jagged Blackwell attention
Meta says its 3.2K-line TLX attention kernel beats FlashAttention-4 on jagged B200 shapes by about 13% forward and 50% backward. Both margins come from GEM's broadcast-query workload in bf16, so another team's gain depends on how closely its shapes match.
Reality
- Evidence45
- Adoption25
- Hype gap+15
- Incentives60
- Confidence50
buildConfirmed2 publishers Andreessen Horowitz puts hyperscaler capex near $780 billion this year and tech at roughly 55% of US capital spending. Its charts also show free cash flow being squeezed, and the hyperscalers are relying more on debt to fund new chip capacity.
Reality
- Evidence55
- Adoption30
- Hype gap+20
- Incentives75
- Confidence60
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
Three consortia will run a free tier with no usage caps on state-supplied hardware, under a rule that Korean models carry at least half the load. The ministry has published neither the cost nor a target for uptake.
Perspective Coverage
5 publishers
- Builder
- Builder 32%
- Operator
- Operator 39%
- Investor
- Investor 29%
Reality
- Evidence60
- Adoption
- Not scored
- Hype gap+40
- Incentives65
- Confidence58
Nscale needs a $590 million second half to hold ByteDance under 20% of 2026 sales
Nscale filed to list on the NYSE at $30bn to $35bn with one unnamed customer behind 73% of its 2025 revenue and 52% of its first half of 2026. Crypto Briefing reports that customer is ByteDance.
Reality
- Evidence40
- Adoption50
- Hype gap+40
- Incentives72
- Confidence45
NVIDIA's Nemotron 3 Super routes each token through a tenth of its 120 billion parameters
NVIDIA says multi-agent systems burn up to 15 times the tokens of a standard chat, and Nemotron 3 Super is its open-weight attempt to make each of those tokens cheaper to produce. The efficiency figures come with NVIDIA's own hardware and its own predecessor as the baselines.
Reality
- Evidence38
- Adoption18
- Hype gap+34
- Incentives88
- Confidence57
buildOne report1 publisher Removing Mamba-2 from a four-mixer stack cost more than making the stack uniform
VIDRAFT AI Research spread seven sequence mixers over 49 layers on a 7x7 Latin square so depth placement could not confound composition. The ablation that carries the finding ran on a 16-layer proxy at 700.9M parameters.
Reality
- Evidence48
- Adoption20
- Hype gap+15
- Incentives40
- Confidence55
Artificial Analysis's Intelligence Index carries a quarter of Korea's sovereign AI score
Korea's Dokpamo program gives the benchmarking firm's index 25 of the 100 points that decide which teams advance. In the results published on August 27, the team that led that component finished last of the four.
Reality
- Evidence30
- Adoption55
- Hype gap+25
- Incentives55
- Confidence38
Seoul is buying distribution rather than compute. The government is handing operators a donated chip grant, has set an 80 percent domestic serving floor, and hasn't published what the programme will cost - all in a country where roughly 1.8 million people pay for any AI at all.
Perspective Coverage
3 publishers
- Builder
- Builder 25%
- Operator
- Operator 37%
- Investor
- Investor 38%
Reality
- Evidence55
- Adoption28
- Hype gap+38
- Incentives58
- Confidence57
buildOne report1 publisher Neurophos schedules a low-rate Austin production line two years before its chip ships
Neurophos has made Kymeta founder Nathan Kundtz its chief technology officer, opened a Sunnyvale office and scheduled a low-rate Austin production line for later this year, with the chip due in 2028 and no benchmarks published.
Reality
- Evidence45
- Adoption6
- Hype gap+42
- Incentives70
- Confidence55
buildOne report1 publisher NVIDIA's 2.5x concurrency figure rests on a 64K prompt reused 76 percent of the time
NVIDIA reports 2.5x more concurrent users for Nemotron 3 Ultra from its packaged NIM serving stack. The gain comes from five interacting layers measured on one agentic traffic shape, and NVIDIA ships the harness to retest it.
Reality
- Evidence55
- Adoption
- Not scored
- Hype gap+18
- Incentives85
- Confidence62
buildOne report1 publisher A 33% throughput gain kept this 3.848B pretraining run under $1,000
Hugo Vergnes reports 0.384 CORE from 65.3 billion tokens on eight rented B200s. The $998 buys 43 hours of node time, not the 139 GPU-hours of failed run that produced the recipe, and not the evenings that wrote the framework.
Reality
- Evidence52
- Adoption12
- Hype gap+22
- Incentives45
- Confidence62
buildOne report1 publisher Identical B200 clusters are repricing 50-60% higher instead of falling, investor says
Gavin Baker's checks say the same hardware costs half again more than it did seven months ago, so the number worth modelling is the price at which end customers finally cover compute, power and financing.
Reality
- Evidence38
- Adoption62
- Hype gap+18
- Incentives72
- Confidence44
Thomson Reuters spent $40 million to own the layer above the open weights
Thomson Reuters puts development of its first proprietary model at roughly $40 million, with under $450,000 of that in the final training run, and says the total still excludes the startup it bought and the cost of keeping the model running.
Reality
- Evidence45
- Adoption20
- Hype gap+20
- Incentives68
- Confidence52
18x per joule in 16 months, and most of it was not your model choice
Stanford's Hazy Research group put a decomposed number on local inference efficiency. Hardware did more of the work than architecture, which changes who captures the gain.
Reality
- Evidence34
- Adoption16
- Hype gap+38
- Incentives
- Not scored
- Confidence30
H100 rentals are back to $2.35 an hour, and your AI cost model is stale
GPU rental prices fell more than 60% in 2025. SemiAnalysis data says H100s have since climbed about 40% to $2.35 per GPU-hour, with 12 to 18 month waits for clusters.
Reality
- Evidence28
- Adoption38
- Hype gap+32
- Incentives55
- Confidence30