01 · Open-weight frontier
Kimi K3 weights drop — the first 3-trillion-parameter model on a free license
Moonshot publishes its 2.8T-parameter Stable LatentMoE under a Modified MIT license today: 1M-token native context, native vision and video, MXFP4 quantization at a 594GB download, and Kimi Delta Attention that runs 6.3× faster at 1M context than K2. The API is priced at $3 input / $15 output per million tokens with a 90%+ cache-hit rate on coding workloads. Independent testers flag a 51% hallucination rate not disclosed in Moonshot's benchmarks — production use should validate before committing.
Read the launch post →
02 · Open-weight frontier
Ant Ling ships a 124B MoE that claims 1T-flagship parity at 1/12 active compute
Ant Group releases Ling-3.0-Flash — 124B total / 5.1B active per token, native 256K context with 1M extensions, and KDA-MLA hybrid attention on a 5:1 layer ratio. The model runs free on OpenRouter and the Vercel AI Gateway through August 3. The release is the second confirmation in 48 hours that small-active MoE with hybrid attention is a multi-lab pattern, not a Moonshot quirk.
Read the model docs →
03 · Infrastructure tooling
Vercel Labs ships TypeScript-to-native compiler Scriptc
Scriptc compiles ordinary TypeScript to native executables via tsc → IR → LLVM, with no Node, V8, or QuickJS embedded unless explicitly opted in. 666 GitHub stars in four days, Apache 2.0, and HN rank #4 with 119 points. The pattern matters for any workload where TS ergonomics were being held back by the cold-start tax of a JS engine — edge functions, serverless handlers, and CLI tools all gain a new default option.
Inspect the project →
04 · Capital stack
$700B of hyperscaler capex meets the cost-of-capital test
Alphabet, Microsoft, Meta, and Amazon are tracking toward roughly $700B of combined 2026 capex and could exceed $1T in 2027. Alphabet raised its 2026 guidance by $15B to $195B–$205B and lost about $293B of market value the day it announced the increase. The signal: investors will keep paying for AI infrastructure, but only when workload-level economics are visible. Token routing, usage observability, and disciplined inference cost are no longer nice-to-haves.
Read the capex analysis →
05 · Capital concentration
Venture capital is concentrating around AI infrastructure bottlenecks
AI represented more than 70% of global venture financing in Q2 2026, and roughly 60% of the quarter's capital went to rounds larger than $1B. About 88% of AI funding went to US companies, while OpenAI and Anthropic together absorbed ~$217B, roughly 43% of first-half venture volume. The implication is straightforward: scarce infrastructure — compute, inference hardware, energy, memory, robotics data, deployment capacity — is where valuation premiums live today.
Read the VC analysis →