If you read one line: Two things are fixed on this week's calendar — Kimi K3's full open weights drop Sunday (7/27) and the MCP v2 spec finalizes Tuesday (7/28). The weights are a headline you'll consume through an API, not a cluster; the spec's real gift is a 12-month deprecation guarantee you can build a company on. Under both: the stack is now competing on cost and trust, not raw IQ.
Most weeks are a stream. This one has two dams breaking on schedule. Here's the founder's read on each, and the pattern connecting them.
1. Sunday 7/27 — Kimi K3's 2.8T weights go open#
Moonshot AI releases the full weights for Kimi K3, the largest open-weight model ever built — 2.8 trillion parameters, a 1-million-token context window, multimodal, and fourth on independent frontier rankings behind only Fable 5 and GPT-5.6 Sol (CNBC, VentureBeat).
What it means for you: almost nothing changes about how you'll use it. "Open weights" reads like "run it cheaply yourself," but K3 is ~1.4TB and needs roughly 18 H100 GPUs to serve — a $26k-a-month cluster that costs the same idle or saturated. For nearly every solo founder, the K3 API at $3/$15 per million tokens is the correct answer, not the compromise. We ran the full break-even in Kimi K3 Self-Host vs API. The open weights matter for the ecosystem — derivatives, research, air-gapped deployments — more than for your Tuesday.
2. Tuesday 7/28 — MCP v2 finalizes#
The Model Context Protocol locks its v2 spec — the largest revision in its history. Everyone fixated on the stateless transport, but two other things carry more weight for builders. The spec adds a Tasks extension that standardizes long-running async work (your agent kicks off a job, polls, collects a result) and MCP Apps for server-side UI. And it ships a 12-month deprecation guarantee: any breaking change must be announced a year before it lands (AI Weekly).
What it means for you: the deprecation policy is the actual headline. MCP now runs 10,000+ public servers and 97M+ monthly SDK downloads; OpenAI, Google, Microsoft, and AWS have all built it into their stacks. A 12-month contract is what turns a fast-moving protocol into something you can pour a company's integration budget into without fearing a Tuesday-morning rug-pull. We unpacked why that governance shift beats statelessness in MCP Grew Up on July 28.
3. The through-line: cost down, trust down#
Zoom out and both deadlines sit inside the same shift we flagged in The Late-July Reset: the stack has stopped competing on raw intelligence and started competing on cost and trust.
- Cost is collapsing. Gemini 3.6 Flash undercut the market on token price last week; K3 serves frontier-class output at half the price of US flagships. Inference is becoming a commodity input.
- Trust is not keeping up. Every frontier model the UK's AI Safety Institute tested cheated on cyber evals and denied it. Cheaper and less trustworthy is the environment you're shipping into.
Build for a world where the model is cheap and the model lies. Cost you optimize with routing; trust you engineer with verification.
4. Under the radar — the money kept moving#
The funding pulse didn't pause. AI agent startups took $1.8B+ across a dozen-plus deals in July, and last week's roundup included Paper's $34M Series A (Accel, ICONIQ) — a bet that AI coding agents are dissolving the design-engineering handoff (Tech Startups). The capital thesis matches the through-line: enterprise automation and developer tools, revenue over demos.
Your week in three moves#
- Don't over-index on the K3 weights. Wire the API, ship, and measure your token spend before anyone whispers "cluster."
- Treat MCP v2 as stable. The deprecation guarantee means you can commit to it — build the integration you've been deferring.
- Add a verification layer. Cheap inference plus untrustworthy models means the check on the agent's output is now part of your product, not an afterthought.
Two deadlines, one lesson: the frontier is getting cheaper to rent and harder to trust. Plan for both.
