Yesterday was geopolitics and GPT-Live. Today is the bill.

Alphabet raised its 2026 capital spending forecast to between 195 billion and 205 billion — up from the prior 180–190 billion range — and the market did not clap. Shares of Google’s parent fell roughly 6% as investors re-ran the free-cash-flow math on Big Tech’s AI buildout. Cloud still beat (Google Cloud revenue growth around 82% in the print), but the narrative flipped from “growth is real” to “how long can this burn before returns show up.”

That is the same tension we have been tracking all summer: agentic products need clusters, clusters need capex, and capex eventually needs a P&L story that is not “trust us, agents are next.”

Etched: inference silicon keeps getting funded

While Alphabet was explaining why it needs another tens of billions of dollars of GPUs and data-center steel, Etched raised about 300 million at a 10.3 billion valuation, per Reuters. The company has been in the “transformer-specialized inference silicon” conversation for months — earlier rounds and contract chatter put them in the multi-billion club with serious booked demand.

The pattern is familiar: frontier labs and hyperscalers spend like utilities building a grid, and a second wave of chip startups races to own the inference side of that grid. If training is where the prestige lives, inference is where the recurring cost lives — and every coding agent you leave running overnight is an inference bill with a persona.

Worth keeping on the hardware shelf next to your usual local GPU builds if you are still self-hosting agent stacks instead of renting someone else’s.

Databricks × Azure Cobalt: agentic workloads get a home

Also today: Databricks expanded its Microsoft Azure partnership through the 2030s, saying it will ramp Azure usage for its own ops and analytics and lean harder on Azure Cobalt — Microsoft’s Arm-based custom processors — for data-intensive and agentic AI workloads (Reuters).

That last phrase matters. “Agentic” is no longer a demo tag on a keynote slide; it is showing up in hyperscaler and data-platform press releases as a named workload class that gets its own silicon story. Cobalt for agents is the quiet admission that long-running tool loops, retrieval, and orchestration look different from one-shot chat completions.

If you are wiring agents into lakehouse stacks, the Azure/Databricks gravity well just got another brick.

Flat-rate coding agents and the cost wall

On the software side of the same cost story, Azumo opened Valkyrie — an open-weight coding agent pitched explicitly around predictable, flat-rate development cost rather than token roulette (MarTech Series). Friends-and-family first, public later. The pitch is simple: engineering teams want agent help without a surprise invoice that looks like a second payroll line.

That is the same pressure we covered when Gartner-style agent spend estimates hit 2K–20K per developer per month. Open weights plus flat pricing is the counter-move to “your agent burned three thousand dollars overnight grepping the monorepo.”

For teams still shopping harnesses, the practical shortlist remains mixed: Claude Code / Codex-class cloud agents when you need frontier, local stacks and open weights when you need budget predictability, and hybrid routing when you are adult enough to meter both.

Operator notes before the weekend

A few non-capex items worth logging while the markets freak out about steel and silicon:

• DeepSeek API aliases — migration pressure around deepseek-chat / deepseek-reasoner continues into the July 24 window. If you still hardcode the old aliases in agent configs, fix it before the break, not after the 500s start paging you. Prefer explicit v4-flash / v4-pro model IDs and check thinking defaults on flash.
• White House 30-day pre-release review chatter keeps circulating (OpenAI / Anthropic / Google in the frame; Meta often described as outside). Treat the “voluntary” label with the same skepticism we used after the Fable/Mythos access drama earlier this summer — the June EO framework already normalized government early look. Enforcement still shows up as access friction, not just press releases.
• Continuity from yesterday’s dispatch: sanctions language on Chinese open models + GPT-Live shipping is still the geopolitics layer under today’s spend story. Capital does not erase policy risk; it just funds the datacenters that policy will eventually regulate.

If you are building agent systems this week, the practical stack question is less “which model wins the leaderboard” and more “who owns my inference bill when the agent runs for seven hours.” Alphabet just told the market the bill is still going up. Etched and Cobalt are bets on making that bill silicon-efficient. Valkyrie-style open-weight flat rates are bets on making it human-budgetable.

Pick your layer. The spaghetti does not get cheaper by ignoring the invoice.

Want this in your inbox every morning? Subscribe to the SpaghettiStories newsletter.

Some links may be affiliate links. If you’re buying hardware to run local models, this affiliate link helps keep the lights on. For agent-friendly keyboards and docks while you supervise long runs, see mechanical keyboards and USB-C docks.