5 min read · 33 sources

The number that matters today is 69%. Datadog reports that in March 2026 customer traces, 69% of all input tokens were system prompts - static, repeated, cacheable text - and it has now shipped prompt caching for Anthropic and OpenAI models to attack exactly that. This is the rare piece of AI-infrastructure news that is simultaneously an engineering fact and a margin statement: every AI-heavy operator is paying frontier-model rates to re-send the same instructions thousands of times, and the tooling that fixes it sits in observability, not in the model providers’ own consoles. That is a genuine second-order read: the model vendors are the obvious names in the AI-cost story, but the beneficiary here is the layer that can see the traces.

When 69% of your tokens never change, caching stops being an optimization and becomes the business model.

Inference economics move from headline to plumbing

The Datadog finding pairs with two quieter data points in today’s material. GitHub Copilot’s cost-efficiency experiments showed that shortening tool responses saved local tokens but increased global token usage and task time - a clean demonstration that local optimization in agent systems is not global optimization, and that naive context-trimming can backfire. And the Pragmatic Engineer piece on Uber’s Minions harness describes companies benchmarking all frontier and open models weekly and caching prompts to save money. Three separate sources converging on the same problem in the same week is what a theme building looks like, not a one-off. The read-through: whoever owns the trace owns the cost lever, which is why the observability names are the non-obvious beneficiaries of the agentic buildout.

Cloudflare's quiet compounding

Two Cloudflare items, neither of which will make a headline but both of which deepen the moat. First, the Automatic Key Exchange work that prefers post-quantum hybrid X25519MLKEM768 cut HelloRetryRequests from ~52% to 3.7% and saved over 150ms at p90 on handshakes for hundreds of thousands of domains - post-quantum migration going from tax to non-event is exactly the kind of thing that keeps enterprises from leaving. Second, the rebuilt workerd module registry now supports every stable Node.js API by default, with apps up to 64 MiB. The second-order effect of the Node compatibility story is that it lowers the exit cost from traditional Node hosting - the displaced parties are the commodity app-host incumbents, not the hyperscalers.

Agents that pay agents

Coinbase’s Bazaar MCP is a discovery marketplace where AI agents can search for, pay for, and use tools independently via x402. Strip the framing and this is settlement infrastructure for machine-to-machine commerce - the missing piece between “agents everywhere” narratives and agents that can actually transact without a human in the loop. It extends a Coinbase thesis already showing up in fee revenue on the tokenized-equities side. Related, from the same agentic strain: East River Source Control as a possible git successor because agentic development strains current version control - a speculative but real signal that the agent era is starting to break tooling designed for human commit cadence.

Stale ground: Oracle, Microsoft-Salesforce, Nvidia-Palantir

Be plain about it: most of the rest of today’s material is continuing coverage. Oracle’s 121% cloud infrastructure growth is real but was already in the archive this week with the capex-stays-flat caveat attached - today’s repeat adds nothing. Microsoft’s AI-powered converter aimed at Salesforce and ERP installs is on its third consecutive day of coverage; we have been negative Microsoft and negative Salesforce on this dynamic since the 9th, and nothing in today’s material changes the read. The Nvidia-Palantir supply-chain pairing is likewise a second-cycle story. The safety cluster - Anthropic’s alignment lead putting >10% on existential risk, the Anthropic biosecurity blocks, the OpenAI board member’s warning, Altman’s openness to slowing frontier development - is a genuinely building narrative, but it cuts both ways for every name simultaneously and produces no differential signal today.

Capacity as a competitive fact

The one item with a clean second-order read in the OpenAI cluster: the pause on new Pro subscriptions due to Astra demand. A capacity constraint at the market-leading voice agent is not just an OpenAI problem - it is a window for Google’s voice stack and for Microsoft’s enterprise distribution to capture workloads that cannot get access. GPT-Live-1’s 30-point gain on Full Duplex Bench and the 80% codebase reduction at one customer suggest the demand is earned, which makes the constraint durable enough to matter for routing decisions.

Overall: a medium day. One genuinely new economic fact (the 69% number), two solid platform wins at Cloudflare, one credible agentic-commerce primitive at Coinbase, and a lot of second-week coverage wearing first-week clothes.

Scoreboard

The book

Trades decided 2026-09-11

Filled at the next session’s open, not today’s close.