Coffee's for Closers

The CFO Finally Read the Bill

"Last year the executives were letting every flower bloom. Now they're coming in with a lawnmower." That's an enterprise exec at this week's Databricks Data + AI Summit, telling The Deep View what happened the day the inference invoice landed. Another one, same room: "The coding agents have run our token bills out of control." These are not dramatic people. Their CFOs are about to be.

Ali Ghodsi said the quiet part into a microphone at his own keynote. The spend is "completely unsustainable," and the number one question he gets now is "how do we curb the cost but still invest in AI."

We wrote this column eight days ago. Buy Wins, Not Players, June 11: the nerds would spend the summer fighting over the last 10% of model quality while the allocators who meter and route quietly took the season. I'm not running a victory lap. I'm pointing at the receipt, because the most expensive lesson in this business is the one you learn a quarter late, and the bill just came due for everyone who waved every flower through.

The mechanism is beautiful if you're standing on the right side of it. Databricks just crossed $6.9 billion in revenue, up 80% in a year. Its lead over Snowflake was $490 million in March. It's $1.6 billion today. Tomasz Tunguz calls companies like it the "first derivative of inference" — the ones sitting in the token path, taking a cut of every query. The same token bill wrecking the buyer is the invoice making the toll-collector rich.

Two choices, only two. Be the toll-collector, or get disciplined about the toll.

Discipline looks like this. Stop firing every query at a frontier model in the cloud. Using Opus to answer a FAQ is cutting a daisy with a chainsaw. Route the 80% of dull, high-volume work to a cheap open model, run it on your own iron if the volume justifies the buildout, and save the frontier for the 20% of high-value work where the extra IQ actually earns its keep. Find your superstars. Bench everyone else. The bench just got absurdly cheap.

You don't have to take my word anymore. Microsoft made the OpenAI model optional in Copilot this week and is testing a fine-tuned DeepSeek for the cheap lane. The largest software company on earth just shipped the routing map as a product feature. That's the whole thesis, boxed and sold.

One warning on where you swing the scalpel. Gartner says half the companies that replaced their support teams with AI will be rehiring humans by 2027 — not because the AI flopped, but because they confused cutting payroll with adding value. Coffee's for closers. That's a rule for your token budget. It's a very expensive rule to point at your people.

At COAI today: the full Signal/Noise — with the routing map and the eight questions that tell you which workloads earn a frontier token and which get routed down — is live at getcoai.com.

Outsider Labs Advisory →

$1.6 billion — the revenue gap between Databricks and Snowflake, up from $490 million in March. One company sits in the token path and clips a fee off every query; the other mostly watches the queries roll past. Same market, same quarter, a billion-dollar swing in ninety days. The token bill wrecking the buyer is the invoice making the toll-collector rich. The only question that matters is which end of that you're standing on.

Build your routing map before your CFO builds it for you. List your ten highest-volume AI workloads and mark which ones actually need frontier IQ. Most won't. The jobs answering FAQs, tagging tickets, and summarizing docs can run on a cheap open model today at a fraction of the cost — and you'll have the answer ready when finance asks why the bill doubled.

Move one workload off the frontier this week and measure the delta. Pick a high-volume, low-stakes job and run it on an open or local model. Microsoft just made the OpenAI model optional in Copilot and is testing a fine-tuned DeepSeek for precisely this lane. If the biggest software company on earth is routing down, your FAQ bot can stop paying frontier prices to cut daisies.

Read the Gartner number before you cut a single head. Half the firms that swapped support staff for AI will rehire by 2027. Route your tokens ruthlessly; route your people carefully. Find the superstars the agents make 10x more productive and protect them. Cutting payroll and adding value are not the same line on the P&L, and the bill for confusing them comes in 2027.

Wednesday, June 17

The token panic finally reached the C-suite. (Full analysis above.) Execs at Databricks' summit told The Deep View inference costs have gone "completely unsustainable." Databricks crossed $6.9B in revenue, up 80%, widening its lead over Snowflake to $1.6B. The pick-and-shovel guys are quietly winning the gold rush.

Microsoft made the OpenAI model optional. (Feeds the analysis above.) Copilot Cowork went generally available worldwide with multi-model support, and Microsoft is testing a fine-tuned DeepSeek V4 for the cheap lane. When your biggest distributor stops treating your model as the default, the model was never the moat.

Washington's kill switch went global. The export order that pulled Claude Fable 5 and Mythos 5 is still dark a week on, and now the G7 is openly talking about it — allies waking up to the fact that US frontier AI can be revoked by a phone call. Anthropic, near a $1 trillion valuation, just learned its best model is a license, not an asset.

SpaceX closed the Cursor deal, and it's a front door for xAI. $60B all-stock, filed Monday. The open question for the 60% of the Fortune 500 running Cursor: do Claude and GPT-5.5 stay first-class in the model picker, or get nudged toward Grok and xAI's Colossus compute? Watch the picker — it's the first real tell.

— Harry and Anthony

Sources: