Uber Cracked the AI Cost Problem. Now Comes the Hard Part.

Start with the question nobody on Wednesday’s earnings call asked directly: if the company that did more than anyone to inflate enterprise AI bills just told investors it figured out how to quadruple AI usage while paying less per token, who absorbs the revenue gap on the other side?

Uber CTO Praveen Neppalli Naga posted that claim on X on August 5, the same afternoon the company reported Q2 results. The timing was not accidental. Naga called it “another signal that we’re coming to the end of the so-called tokenmaxxing era.” The company that lit the fuse on corporate AI overconsumption is now the one calling time on it. That is worth paying attention to.

How the Budget Collapsed

Uber exhausted its entire 2026 AI budget by April. Four months into the calendar year, Anthropic’s Claude Code had spread across roughly 5,000 engineers faster than the company’s finance models anticipated. Naga confirmed the overrun in an interview with The Information, saying he went “back to the drawing board” on the company’s spending assumptions.

The mechanism behind the blowout was culture as much as code. Uber encouraged employees to use Claude Code as much as possible, going so far as to build leaderboards ranking engineers on usage. Monthly cost per engineer ran $150 to $250 on average. Power users landed between $500 and $2,000. Naga himself reported spending $1,200 in a two-hour personal demo session. The company had gamified its way into a budget crisis.

By May, the COO was publicly questioning the whole exercise. Andrew Macdonald said it was hard to connect rising Claude Code usage to anything consumers actually saw. “That link is not there yet,” he said. “Maybe implicitly there’s more that is getting shipped, but it’s very hard to draw a line between one of those stats and ‘Okay now we’re actually producing like 25% more useful consumer features.'” By June, Uber installed a $1,500 monthly cap per employee across its agentic coding tools, including Claude Code and Cursor.

That three-act sequence, from unlimited encouragement to public skepticism to hard caps, played out inside six months. Uber was not alone. The rest of the industry was working through the same reckoning, asking what all those tokens actually bought.

The Reversal the CTO Is Now Claiming

Eight months into the year, Naga is describing a different company. The count of Uber employees using frontier AI tools has quadrupled since January. AI costs are declining. The mechanism, he wrote, is not rationing: Uber improved prompt caching, adjusted its default model selections, ran trials with open-weight models, and gave engineers real-time visibility into their usage and hourly cost.

“You might expect costs to rise as adoption accelerates,” Naga wrote. “We’ve seen the opposite. Not because we’ve restricted access, but because we’ve treated efficiency as an engineering problem rather than a budget problem.”

CFO Balaji Krishnamurthy added that the company is seeing doubled code output per engineer. That is a productivity claim. It is not a product revenue claim. The COO’s May skepticism about whether higher AI output translates into consumer features has not been formally walked back. Both things can be true at once: engineers are faster, and customers may not yet feel it.

The Business Behind the Stock

Set the AI debate aside for a moment. Uber’s Q2 numbers were genuinely strong in the places that matter. Revenue came in at $14.19 billion, up 12% year-over-year, against expectations of $14.24 billion. Net income reached $2.39 billion, or $1.17 a share, compared to $1.35 billion, or 63 cents a share, a year earlier.

Gross bookings rose 22% year-over-year to more than $58 billion, above the top end of guidance. Trailing 12-month free cash flow crossed $10 billion for the first time. Mobility gross bookings reached $28.99 billion, up 22%. Delivery bookings jumped 26% to $27.46 billion. Trips rose 18% to 3.87 billion. Monthly active platform consumers grew 16% to 208 million.

The stock fell anyway. Shares closed 5.3% lower on Wednesday, as investors locked onto Q3 guidance that came in below expectations. Uber’s 52-week high was $101.99. At roughly $75, the stock sits about 26% below that level. The market is pricing Uber as a ride-hailing company with an R&D cost problem, not as a platform posting record free cash flow and, if the CTO is to be believed, winning the AI efficiency race among large enterprises.

Corporate G&A and Platform R&D costs increased 18% in the quarter. The company spent $951 million on R&D in Q1 2026 alone, nearly 17% more than a year earlier. If Naga’s efficiency claims hold through Q3, that trajectory should bend. A deceleration in R&D growth while headcount holds steady would be a meaningful margin story heading into 2027.

The Second-Order Problem: Anthropic’s IPO Math

Here is where the Uber story becomes an investment question that reaches well beyond UBER shares. Uber burned through its Claude Code budget in four months and then responded by engineering its way to lower cost-per-token. It is not alone. Companies across the industry have begun placing engineers on AI budgets as tokenmaxxing bills arrive, a sign the free-for-all is giving way to spreadsheets.

Microsoft was reported earlier this year to have begun canceling direct Claude Code licenses and routing engineers back to GitHub Copilot, citing monthly bills running $500 to $2,000 per engineer. That is a named enterprise account walking away from frontier token spend toward a cheaper alternative. It is one data point, but it is a named one.

The timing of this efficiency shift collides directly with Anthropic’s IPO ambitions. Anthropic’s annualized revenue run rate crossed $47 billion in May 2026, up from $9 billion in January, a pace with almost no precedent in enterprise software. The company closed a $65 billion Series H at a $965 billion post-money valuation and filed a confidential S-1 with the SEC on June 1, targeting an October Nasdaq listing.

Claude Code alone contributes an estimated $2.5 billion of that run rate, according to analysis published ahead of the IPO. Coding workloads are the highest-value token category Anthropic has: sticky, output-heavy, and consumed by users who can generate more tokens in a single agentic session than a casual chatbot user does in a month. Enterprise coding customers are disproportionately important to Anthropic’s revenue because they are the ones running sessions at $1,200 a pop.

Uber’s efficiency shift is exactly the behavior that compresses that model. More engineers using Claude Code at lower average cost per token means flat or declining token revenue even as enterprise adoption broadens. Uber’s engineers did not stop wanting to use the product. They ran out of budget to pay for it at full consumption. That is a demand problem wearing the mask of a satisfaction problem. The open question is whether falling token prices can re-expand the demand curve fast enough to offset what enterprises are optimizing away.

The Risks

Uber’s efficiency reversal is self-reported and based on roughly three months of data. Naga has every incentive to declare the problem solved on earnings day. The productivity claims, doubled code output, faster deployment cycles, have not been independently verified. The COO’s May skepticism about consumer-facing improvements has not been formally withdrawn. Efficiency and output are not the same thing as product.

On costs, Corporate G&A and Platform R&D grew 18% against a quarter where revenue grew 12%. That gap does not close on a CTO blog post. If R&D expenses continue outpacing segment profit growth, operating leverage remains constrained regardless of what the token cost per hour is doing.

The broader cultural reversal also introduces its own execution risk. Tokenmaxxing worked, in part, because engineers believed they had something close to unlimited access. Caps change behavior in ways that are hard to predict. Some engineers will optimize. Others will simply do less. The difference between those two outcomes matters for whether the productivity claims survive into Q3.

What Investors Should Watch Next

Three numbers will tell investors whether August 5 was a turning point or a well-timed announcement. First, the Q3 R&D expense line in November. If Platform R&D growth decelerates from 18% while headcount holds flat, the efficiency claim has a foundation. Second, Q3 operating margin. Uber’s guidance implied an 18-basis-point expansion over Q2. Whether AI cost savings flow to margins or get reinvested elsewhere is the variable that separates a good story from a good investment. Third, any disclosed metric on autonomous vehicle unit economics. Robotaxi deployment in partnership with Waymo is the highest-stakes R&D line in Uber’s portfolio and the one where token efficiency arguments matter least.

Watch Anthropic’s disclosures just as closely. The company is targeting an October listing. Prediction markets currently put the probability of a 2026 IPO at roughly 76%. The public S-1 prospectus, when it arrives, will need to address enterprise churn and average revenue per customer directly. If the largest, most visible enterprise AI spender in the country is optimizing costs downward at the same time Anthropic’s roadshow kicks off, the S-1 will need a compelling answer for how a $965 billion valuation holds when best customers are actively shrinking their invoices.

Bottom Line

Uber’s tokenmaxxing arc has always carried two audiences: investors watching the R&D line in Uber’s income statement, and investors in the broader AI infrastructure trade trying to gauge whether enterprise token demand is durable at current prices. On August 5, the CTO gave both a data point neither had before. A company that burned its full 2026 AI budget in four months is now claiming it quadrupled adoption while cutting cost-per-token, without restricting access.

If that claim holds, the Uber bull case sharpens considerably: record free cash flow, margin expansion on deck, a 26% discount to the 52-week high, and an R&D cost structure that could finally bend in the right direction. The bear case is more straightforward. Q3 guidance missed. R&D costs are still growing faster than revenue. The COO’s skepticism about consumer-feature output has not been resolved. The autonomous vehicle transition remains expensive and long-dated. The CTO’s efficiency reversal is a reason to watch Uber’s November numbers with fresh attention. Whether it is a reason to own the stock at $75 depends entirely on whether the Q3 data shows up to match the August story.