Citigroup says 14,000 engineers on Devin cut vendor spend 30%; Cognition's $48 billion mark waits on December
Jane Fraser's disclosure is Devin's first bank-scale usage number, funded by other vendors' shrinking budgets. Cognition's $48 billion round now waits on a December run rate of $4 to $5 billion that exists only as a forecast.
Vincent Jiang · 3 min read
Fourteen thousand engineers, and the vendor bill behind them
On Citigroup's third-quarter call on October 13, 2026, chief executive Jane Fraser gave Devin its first bank-scale number: about 14,000 of the bank's engineers use Cognition's coding agent, products are launching 30 to 40 percent faster, and "we've reduced our tech vendor spend by 30% year-over-year as a consequence, and that's still got a long way to run" 123. Citi+1.28% — Citi, up 1.28 percent today is re-engineering about 100 manual processes with AI and pulling contracted work back in-house 1. The causal chain is the chief executive's own, and no bank disclosure separates what Citi pays Devin from what it saves. The disclosure still matters: an agent becomes a budget line by shrinking other vendors' lines.
The $48 billion is priced on December
Cognition, founder Scott Wu's applied AI lab, raised more than $2 billion at $48 billion on September 8, 2026, four months after a $26 billion round, on run-rate revenue it says rose from $492 million in May to almost $900 million 4. It passed $1 billion on September 25, a company-reported figure 5. The Information expects $4 billion to $5 billion annualized by year-end 4. Cursor, the comparable exit, sold to SpaceX+0.20% — SpaceX, up 0.20 percent today at $60 billion on annualized revenue that had surpassed $2 billion: under 30 times, against about 48 times the $1 billion print for Cognition's September mark 4.
Cognition's stated run rate doubled twice since May; December is a forecast, not a print
- Stated run rate
- Estimate
Data
| Stated run rate | |
|---|---|
| May '26 | $0.49 |
| 8 Sep '26 | $0.9 |
| 25 Sep '26 | $1 |
| Dec '26 (est.) (estimate) | $4.5 ($4–5) |
Rationing by output, while the rent runs
October's other Cognition news runs the other way. Higher Devin usage limits now depend on estimated engineering hours returned per dollar of session cost, validated at a 0.74 log-scale correlation across 233 held-out sessions 6. The Nvidia+0.32% — Nvidia, up 0.32 percent today cluster Cognition rents from CoreWeave+4.40% — CoreWeave, up 4.40 percent today, the Nasdaq-listed AI cloud, costs hundreds of millions a year, and 2026 burn could reach $800 million 4. OpenAI's $20 ChatGPT login now meters Devin's OpenAI calls, with nothing disclosed between the companies 7. The landlord is repricing too: CoreWeave bonds issued in June at 9.675 percent traded near 15 percent yields in the week to October 15, and the shares fell 15.1 percent, even as Cognition's own engineers measured 4.8 times the token throughput of a GB200 baseline running the first production workloads on Nvidia's Vera Rubin racks 8910.
Two readings decide it
The bull case is arithmetic: December at $4 billion to $5 billion compresses the $48 billion to 10 or 12 times revenue, below what Cursor$30B — Cursor, private, latest valuation $30B carried, though that forecast is The Information's, not Cognition guidance 4. Watch the December print, and the first enterprise contract Chris Degnan, Cognition's new chief revenue officer from Snowflake, signs without a consultancy on it; Cognizant+1.30% — Cognizant, up 1.30 percent today has trained more than 3,000 associates on Devin and owns most delivery today 11. The first receipt came out of a bank's vendor budget. The verdict comes from Cognition's own December number.
Deepdive
AI-generated from this story and its cited sources. Not investment advice.
