The Enterprise AI Reckoning Has Arrived

The Big Question

The central debate inside every serious investment committee meeting this week is not whether artificial intelligence is real. It is whether enterprise token demand is durable at the prices that underpin today’s AI valuations.

On August 5, Uber CTO Praveen Neppalli Naga handed that debate a data point it had been missing. Uber, the company that did more than any other large enterprise to inflate AI token bills, publicly declared that it has quadrupled AI tool adoption since January while simultaneously cutting cost-per-token. It called the tokenmaxxing era over. The company that lit the fuse is now calling time.

That claim sits at the intersection of two of the largest investment questions in the market right now: what happens to Uber’s margin structure if it is true, and what happens to Anthropic’s October IPO if every enterprise is doing the same thing quietly.

Why Wall Street Cares

The AI trade has rested on a specific assumption: that enterprise adoption would drive sustained, compounding growth in token consumption. That assumption justified the infrastructure spending arms race, the frontier model valuations, and the GPU supercycle thesis. Anthropic filed a confidential S-1 on June 1, targeting an October Nasdaq listing at a $965 billion post-money valuation on a run-rate of $47 billion in annualized revenue reached as recently as May. Eight of the Fortune 10 are customers. Enterprise clients account for roughly 80% of that revenue, with over 1,000 companies now spending more than $1 million per year.

The entire valuation framework depends on those enterprise customers continuing to consume at scale, or consuming more. Uber’s announcement introduces a third possibility: consuming more users, for fewer dollars. That is the scenario institutional investors are now trying to price.

Separately, Uber itself trades at roughly $71, nearly 30% below its 52-week high of $101.99, despite posting record trailing 12-month free cash flow of more than $10 billion. Analysts at DA Davidson, Needham, and Bernstein all cut price targets after Q2 but maintained Buy or Outperform ratings, with consensus targets clustering between $95 and $109. The gap between what the business is doing and what the stock is pricing is the second reason investment committees are paying attention.

The Bull Case

The optimistic read on the Uber efficiency story is straightforward. A company that was spending irrationally has learned to spend rationally. The tools work. The engineers are faster. CFO Balaji Krishnamurthy confirmed on the Q2 earnings call that the company is seeing doubled code output per engineer. CEO Dara Khosrowshahi reported that autonomous agents built roughly 10% of committed code in Q2. Productivity is real and measurable, even if the translation to consumer features remains contested.

If Platform R&D costs decelerate from their current 18% growth pace while revenue continues compounding at 12%, the margin expansion that analysts have been projecting for 2027 starts to look achievable. Trailing free cash flow already topped $10 billion for the first time. Gross bookings rose 22% year-over-year to more than $58 billion, above guidance. Net income reached $2.39 billion in Q2, up 77% from a year earlier. The stock’s discount to intrinsic value, on those numbers, is significant.

On the broader AI demand question, Goldman Sachs equity analyst Jim Schneider has forecast that token consumption will multiply 24 times between 2026 and 2030, reaching 120 quadrillion tokens per month, as agentic AI applications scale. Under that framework, enterprise efficiency gains compress revenue-per-token but total volume more than compensates. Uber spending less per token while running four times as many users is, in that reading, exactly how a healthy demand curve is supposed to work.

The Bear Case

The skeptical case starts with the same data and reaches a different conclusion. Uber burned its entire 2026 AI budget inside four months. By May, the COO was publicly struggling to connect rising token spend to consumer outcomes. “That link is not there yet,” Andrew Macdonald said. By June, the company had installed $1,500 monthly per-employee caps across all agentic coding tools. The efficiency reversal the CTO announced in August is, in the bear reading, a managed retreat dressed as a strategic breakthrough.

The broader enterprise AI data supports the skepticism. Deloitte’s 2026 survey found that 66% of enterprise leaders report productivity gains, but only 20% see AI-driven revenue growth. Fewer than one-third of corporate decision-makers in a Gartner survey could identify specific financial outcomes from their AI investments. Forrester found that enterprises are deferring roughly 25% of planned AI spend to 2027. The productivity claims are real but diffuse, and diffuse does not show up in a quarterly income statement.

For Anthropic specifically, the bear case is structural. Around 72% of its revenue comes from pay-per-token API calls. Enterprise coding workloads generate the highest token volumes of any use case. A single engineer in an agentic session consumes more tokens in a day than a casual chatbot user does in a month. When Uber optimizes those sessions and Microsoft routes engineers back from Claude Code to GitHub Copilot, the revenue impact is not marginal. It is concentrated exactly where Anthropic’s economics are most exposed.

The Evidence

Uber’s AI budget collapse followed a specific mechanism. Claude Code adoption at Uber rose from 32% to 84% of engineers after the company built internal leaderboards ranking usage. Monthly cost per engineer averaged $150 to $250, with power users running $500 to $2,000. Naga spent $1,200 in a two-hour session during a personal demo. By April, the full-year 2026 AI budget was gone. By June, hard caps arrived.

The reversal Naga described in August involved prompt caching improvements, default model adjustments, open-weight model trials, and real-time cost visibility for engineers. Usage quadrupled. Token costs fell. Whether that is genuine engineering efficiency or the natural consequence of hard caps plus engineers gaming their hourly dashboards is the question no one has cleanly answered.

Microsoft’s behavior adds a named data point on the other side. The company began canceling direct Claude Code licenses in mid-May, routing engineers across its Experiences and Devices division back to GitHub Copilot by the end of June. Monthly per-engineer bills of $500 to $2,000 were the cited reason. That is the largest enterprise software company in the world explicitly choosing a cheaper alternative at scale. It does not affect Microsoft’s broader Anthropic Foundry deal, but it demonstrates that the highest-consumption use case for Claude is also the one most vulnerable to substitution.

Token prices themselves are moving. A Forbes analysis published July 28 documented a rapid collapse in token prices as enterprise spending cuts intensify, with companies increasingly routing routine tasks to cheaper models. Chinese providers including DeepSeek, Qwen, and Kimi captured 46% market share by mid-2026 on price. The direction is set. Flat-rate inference for unbounded agentic workloads was never going to survive contact with a CFO’s quarterly review.

The Mavens’ View

The investors who have thought most carefully about this problem are not debating whether enterprise AI spend will grow. Nearly all of them believe it will. The debate is about the shape of that growth and who captures it.

The more sophisticated framing, coming out of discussions at Goldman Sachs and in Anthropic’s own pre-IPO disclosures, is that falling token prices are not a revenue problem if volume grows fast enough. Agentic AI, which Jensen Huang has argued requires 10 times the compute of standard generative AI, is still in early deployment. The productivity gains that enterprise customers are currently measuring at the task level have not yet translated into the full workflow automation that makes compute consumption genuinely elastic. That transition, from AI-assisted engineers to AI-autonomous workflows, is the demand catalyst that justifies current infrastructure investment.

The countervailing view, reflected in Forrester’s deferral data and in Uber’s COO’s public skepticism, is that the timeline for that transition is being systematically underestimated. Enterprises that cannot demonstrate P&L outcomes from current AI spend are not going to increase that spend ahead of demonstrated returns. The CFO scrutiny of 2026 is the market doing what it should do: requiring evidence before authorizing renewal. Companies that deliver it will expand. Companies that cannot will optimize or cut.

What the mavens agree on is that the easy part of the enterprise AI trade is over. The first wave, selling access to frontier models to large enterprises with unlimited budgets and performance review incentives, has run its course. The second wave requires demonstrating that the spend produces outcomes that justify renewal at the same or higher price points.

What Investors Are Missing

The conversation has focused almost entirely on whether Uber’s efficiency story is real and what it means for Anthropic’s revenue line. The less-discussed implication runs in the other direction: if enterprises are genuinely figuring out how to do more with less, the companies that built infrastructure assuming consumption-based growth would compound indefinitely are the ones with the exposure, not the ones that learned to optimize.

Anthropic’s S-1, when it becomes public, will carry a revenue accounting question that institutional investors are already flagging. Research firm Sacra and analysts at multiple banks have noted that Anthropic reports revenue from cloud resellers on a gross basis, counting total end-customer spend and booking partner payouts as expenses. That practice inflates the headline $47 billion ARR figure relative to net-reporting peers. The actual net revenue, after payments back to AWS, Google, and Microsoft, is materially lower. At a $965 billion valuation priced on that headline number, the gap matters.

The second overlooked implication involves Uber’s autonomous vehicle strategy. Robotaxi deployment with Waymo is the highest-stakes R&D investment in Uber’s portfolio, and it is the one where token efficiency arguments are least relevant. Autonomous vehicle trips on Uber’s platform grew more than 10 times year-over-year, and management is targeting 15 AV cities globally by year-end. That story has been almost entirely ignored in the post-earnings coverage, which fixated on the guidance miss. If the AV unit economics start disclosing at scale, the valuation conversation shifts substantially.

Stocks to Watch

Uber Technologies (UBER). The most directly affected name in both directions. At roughly $71 with consensus analyst targets clustering near $100, the stock reflects a market that is pricing the guidance miss and the R&D cost overhang while ignoring the record free cash flow, the 22% bookings growth, and the AV optionality. The Q3 R&D expense line in November is the binary. If Platform R&D growth decelerates from 18% while headcount holds, the margin expansion case is intact. If it does not, the CTO’s August claims were noise.

Anthropic (pre-IPO, targeting October Nasdaq listing). The highest-stakes valuation question in the AI infrastructure trade. The $965 billion post-money price implies roughly 21 times forward revenue at the current $47 billion ARR run rate, which itself includes gross-basis reporting that inflates top-line figures. Claude Code’s $2.5 billion contribution to that run rate is exactly the revenue stream most exposed to enterprise efficiency optimization. The public S-1 prospectus will need to address customer concentration, net versus gross revenue, and churn among high-consumption enterprise accounts. The roadshow will be a stress test of whether the AI infrastructure premium survives contact with institutional due diligence.

Nvidia (NVDA). The apparent paradox in this debate is that token price deflation and enterprise efficiency gains are, in Goldman Sachs’s analysis, bullish for GPU demand over a multi-year horizon. Volume growth in agentic workloads is expected to dwarf the per-token cost declines. Chinese model providers gaining share on price still run on GPU infrastructure. Falling inference costs historically expand the addressable market rather than shrink it. Nvidia is the name that benefits whether the efficiency story accelerates adoption or whether the agentic transition drives the next volume wave.

GitHub (owned by Microsoft). The clearest quiet beneficiary of the tokenmaxxing reversal. Microsoft’s decision to cancel Claude Code licenses and route engineers back to Copilot at GitHub converted a cost problem into a product consolidation. GitHub is now shifting all Copilot plans to usage-based billing through AI Credits starting June 1, replacing the flat-rate licensing that obscured consumption. That pricing shift gives Microsoft the same consumption upside that Anthropic had while offering enterprise finance teams the predictability they want. If enterprise AI coding spend rationalizes toward fewer, cheaper tools, GitHub Copilot is structurally advantaged.

Cursor (private, $4 billion ARR). The overlooked competitive threat in the coding AI category. Cursor launched its own model in November 2025 and is now running near $4 billion ARR with a 60% enterprise mix. It is disintermediating both Claude Code and GitHub Copilot from the bottom up, building directly on top of foundation models rather than paying frontier model API rates. If enterprise coding spend continues to rationalize, Cursor is the name that benefits from the migration to efficiency-first tooling without carrying Anthropic’s infrastructure cost base.

Bottom Line

The tokenmaxxing era produced the fastest enterprise software revenue ramp in recorded history. Anthropic went from $1 billion to $47 billion in annualized revenue in roughly 18 months. That pace rewarded everyone who bought the infrastructure thesis early. The question August 5 raised is whether the next 18 months look like the last 18, or whether the enterprise efficiency turn Uber described is the leading edge of a broader repricing.

The honest answer is that the data is genuinely mixed. Uber’s productivity claims are self-reported. Anthropic’s revenue trajectory remains steep even as enterprise customers optimize. Goldman Sachs’s agentic volume forecast, if correct, makes current token prices irrelevant within three years. And Uber’s AV business, which is growing 10 times year-over-year in trips and has barely registered in investor coverage this week, could make the entire AI cost debate a footnote in the longer story.

What investment committees should be sitting with is this: the easy consensus trade in enterprise AI was long the model providers and long the infrastructure. That trade has not broken. But it has gotten harder to hold without a view on how enterprises respond when the budget runs out and the CFO asks for outcomes. Uber gave the first detailed public answer on August 5. The rest of the market will spend the next two quarters deciding whether it likes what it heard.