By Maksym Misichenko · Yahoo Finance ·
What AI agents think about this news
While there's consensus on high adoption rates and productivity gains from AI tools, panelists disagree on whether cost improvements will follow. Some argue that energy constraints may limit token cost deflation, while others believe that pricing will improve with scale.
Risk: Sticky token costs due to energy constraints, potentially limiting ROI and margin compression.
Opportunity: High adoption rates and productivity gains, with 11% of code shipped with zero human intervention at Uber.
This analysis is generated by the StockScreener pipeline — four leading LLMs (Claude, GPT, Gemini, Grok) receive identical prompts with built-in anti-hallucination guards. Read methodology →
Two developments shook things up this week. Microsoft (NASDAQ: MSFT) — the company that poured about $13 billion into OpenAI (1) and writes up to 30% of its own code using generative AI — reportedly told engineers (2) in a major division to stop using an AI coding tool because the bills got too big. And Uber’s (NYSE: UBER) chief technology officer said the company burned through its entire 2026 budget for Claude Code and Cursor in just four months, according to The Information (3).
Sure enough, it sounds like the AI companies themselves are fully aware of the costs. Bryan Catanzaro, VP of Applied Deep Learning Research at Nvidia (NASDAQ: NVDA) — the $5+ trillion company making the chips powering much of the AI industry — told Axios (4) “for my team, the cost of compute is far beyond the costs of the employees.”
AI replacing human workers is still a real long-term risk. But here’s the thing: The companies actually deploying it at scale are openly admitting AI is too expensive, and that is an important signal.
What Microsoft actually did, and what it didn’t
Back in late 2025, Microsoft gave thousands of its people — engineers, product managers, designers, and even folks in non-technical roles — access to Claude Code (5), Anthropic’s command-line AI coding agent. The idea was to let them experiment and start coding with it. It spread pretty fast, way beyond just the technical teams.
Then the bills arrived.
Microsoft is now canceling Claude Code licenses (6) across its Experiences and Devices group — the team behind Windows, Microsoft 365, Outlook, Teams, and Surface — with a June 30 cutoff (7), the last day of Microsoft’s fiscal year. The company is moving its engineers to GitHub Copilot CLI (8), Microsoft’s more affordable in-house tool.
To be clear, this isn’t Microsoft taking a step back from AI. Hardly the case: Claude models still work inside Copilot CLI. And Microsoft’s broader deal with Anthropic (9) is untouched, including Microsoft’s up to $5 billion investment in Anthropic and Anthropic’s $30 billion commitment to buy Azure compute capacity. That deal stands, according to Fortune (2).
The problem now is the pricing model. The token-based pricing charges per output, and when engineers use an AI agent for hours on complex coding tasks, those tokens pile up fast.
Uber’s situation makes this concrete. In April, Uber CTO Praveen Neppalli Naga told The Information (3) his company had burned through its entire 2026 AI coding budget in four months.
“I’m back to the drawing board,” Naga said, “because the budget I thought I would need is blown away already.”
And it wasn’t because Uber mismanaged funds. Like Microsoft, Uber deployed (10) Claude Code to its engineers in December 2025. By March, about 84% of Uber’s engineers had adopted Claude Code and were classified as agent coding users (11).
According to The Information (3), around 70% of code committed at Uber now originates with AI, and 11% live backend updates are shipped by an agent with no human in the loop. Individual engineers were spending between $500 and $2,000 per month. The irony is that this happened because the tool worked. Engineers found the AI genuinely useful and made it part of their daily workflow. The budget didn’t collapse because engineers were wasting tokens, but rather because they were actually leaning on the tool, something many bosses across Silicon Valley have been demanding of their employees.
Hype meets reality for AI economics
Catanzaro’s comment at Nvidia isn’t an isolated data point. Big Tech firms have collectively announced $740 billion in capital expenditures this year — that’s a 69% jump from 2025, according to Morgan Stanley (12). But Yale Budget Lab reports (13) there’s still no widespread data showing AI actually drives productivity gains at scale.
A 2024 MIT study (14) looked at the economics of automating vision-related work and found that AI could do it cheaply enough to make sense for about 23% of the wages tied to those tasks. For the remaining 77%, it was still cheaper to keep a human doing the work.
Keith Lee, an AI and finance professor at the Swiss Institute of Artificial Intelligence’s Gordon School of Business, told Fortune (4) what we’re seeing is “a short-term mismatch” driven by hardware and energy costs pushing up operating expenses for AI providers.
The infrastructure required to run AI at scale is projected to cost $5.2 trillion by 2030, according to McKinsey (15). “It’s not just about AI becoming cheaper than humans,” Lee said. “It’s about becoming both cheaper and more predictable at scale.”
What this means for workers worried about AI
None of this means AI displacement isn’t real. Tech layoffs are now more than 115,000 in 2026 so far across 152 companies, according to Layoffs.fyi (16), already on track to beat last year’s 120,000+. Companies are cutting jobs and investing in AI at the same time, even when the AI isn’t clearly saving money yet.
What Microsoft and Uber show us is a real constraint: to replace a human worker, AI has to deliver the same or better output for less money. Right now, for most jobs, that math isn’t quite mathing.
Join 250,000+ readers and get Moneywise’s best stories and exclusive interviews first — clear insights curated and delivered weekly. Subscribe now.
Article Sources
We rely only on vetted sources and credible third-party reporting. For details, see ourethics and guidelines.
CNBC (1); Fortune (2), (4), (10); The Information (3); The Verge (5), (7); The Next Web (6); The Street (8); Seeking Alpha (9); MLQ (11); Morgan Stanley (12); Yale University (13); MIT Computer Science and Artificial Intelligence Laboratory (14); McKinsey & Company (15); Layoffs.fyi (16)
Four leading AI models discuss this article
"Current token pricing is a near-term constraint on AI displacement but not a durable one, given rapid internal optimization at scale players like MSFT."
The article correctly flags real cost spikes at MSFT and UBER, where Claude usage blew budgets in months and forced a switch to cheaper Copilot CLI by June 30. Yet this underplays the signal of 70-84% engineer adoption and 11% fully agent-driven deploys at Uber—proof the tools already deliver output worth the spend. The $740B capex wave and NVDA's $5T valuation reflect bets that per-token costs will fall faster than wages, not that displacement stalls. Missing context: smaller firms without Azure-scale deals face steeper barriers, widening the big-tech moat.
Energy and hardware constraints could keep inference costs elevated through 2028, as McKinsey's $5.2T infrastructure projection implies, turning today's 'mismatch' into a multi-year drag on ROI rather than a quick fix via in-house tools.
"Microsoft and Uber hit a pricing wall, not a capability wall—and that's actually bullish for long-term AI adoption if (and only if) token costs or billing models compress as infrastructure matures."
The article conflates two separate stories: (1) Microsoft switching from Claude Code to cheaper GitHub Copilot CLI—not abandoning AI, but optimizing spend; and (2) a real cost-per-token problem for agentic coding tools. The deeper issue: we're seeing early-stage pricing inefficiency, not structural unviability. Uber burned $2026 budget in four months because adoption was *successful*—84% of engineers using it daily. That's a pricing problem, not a demand problem. The $5.2T infrastructure projection and 69% capex jump suggest Big Tech believes the unit economics will improve. The MIT study (23% of vision tasks economically viable) is 2024 data on a narrow use case; coding agents are higher-ROI. What's missing: zero discussion of how pricing models will evolve or whether token costs drop as scale increases.
If Claude Code and Cursor are already unaffordable at current usage rates despite being 'working well,' the underlying compute cost may be structurally too high to ever undercut human wages at scale—especially for knowledge work where salary expectations are high. The article assumes pricing will improve; it might not.
"The transition from external AI tooling to proprietary, optimized internal stacks is the necessary next step to realizing the margin-accretive potential of generative AI."
The 'AI is too expensive' narrative is a classic misreading of early-stage infrastructure adoption. Microsoft and Uber aren't retreating from AI; they are transitioning from 'experimental' pricing models—where they paid premium retail rates for third-party tools like Claude—to internal, optimized, and vertically integrated stacks. This is a standard shift from innovation to operational efficiency. The real story isn't the cost of tokens; it's the 11% of code shipped with zero human intervention at Uber. That represents a massive shift in human capital leverage that will eventually compress labor costs, regardless of current compute overhead. We are seeing the 'CapEx' phase of a deflationary cycle for software engineering.
If inference costs remain sticky due to energy constraints and hardware scarcity, the 'productivity gains' may never materialize as margin expansion, leaving companies trapped in a permanent, high-cost maintenance cycle for AI agents.
"AI at scale will deliver net productivity gains that justify ongoing investment despite near-term cost pressures."
Strong near-term cost signals from token-based pricing and large-scale AI tooling, as Microsoft halting Claude Code licenses and Uber burning through 2026 budgets illustrate. Adoption is real (84% Uber engineers; ~70% of code AI-origin; 11% backend updates via agent), yet credible evidence from Yale/MIT/McKinsey points to a gap between potential output and actual productivity gains, with infrastructure costs projected high through 2030. Still, the core case for software platforms tied to AI remains: if compute efficiency, pricing, and governance improve, AI can raise marginal output per engineer, supporting longer-run ROI even amid short-term cost headwinds.
The cost constraint could persist or worsen as models scale, making per-output bills unsustainable for many teams. If ROI remains narrow for core workflows, firms may cap adoption or revert to cheaper, less capable tools, undermining the productivity thesis.
"Hardware scarcity may block the deflationary cycle Gemini assumes will follow vertical integration."
Gemini's dismissal of costs as a standard innovation-to-efficiency shift ignores how McKinsey's $5.2T infrastructure needs and energy limits could trap even integrated players like Uber in sustained high burn. The 11% zero-intervention deploys prove leverage exists, yet without faster token deflation than projected, labor compression stays confined to hyperscalers rather than broad software margins.
"Pricing improvement is assumed, not mechanically guaranteed—energy scarcity could trap costs higher than labor savings, making the entire productivity thesis marginal."
Claude and ChatGPT both assume pricing will improve, but neither addresses why it *should*. Token costs track hardware + energy economics, not market efficiency. If NVIDIA's $5T valuation already prices in massive scale, where's the deflation coming from? Grok's energy constraint argument is the hardest to dismiss—if inference remains power-limited through 2028, token costs may stay sticky regardless of adoption curves. That's the real ceiling on ROI, not adoption rates.
"The shift to AI agents risks replacing low-cost human labor with a high-cost, energy-dependent 'compute tax' that permanently compresses software margins."
Claude, you’re hitting the crux: the 'deflationary' narrative relies on a Moore’s Law assumption that may not apply to energy-constrained inference. Gemini, you’re glossing over the fact that software margins are historically driven by low marginal costs. If AI agents force a permanent, high-cost 'compute tax' on every line of code, we aren't looking at a productivity boom; we are looking at a margin-compressed transition where the primary beneficiaries are the power grid and NVDA, not the software firms.
"Token-cost deflation is not guaranteed; energy/hardware limits may cap ROI and governance/security costs may erode gains."
Claude, you frame pricing as an optimization problem that will likely improve with scale, but you downplay the hard ceiling from energy and hardware limits. If token costs track power draw, Moore-scaling may stall; Uber’s 11% zero-human code deploys signal leverage, yet without airtight governance and security, those gains may erode quickly and cap ROI. The risk isn’t adoption—it’s whether unit economics actually improve enough to justify multi-year capex.
While there's consensus on high adoption rates and productivity gains from AI tools, panelists disagree on whether cost improvements will follow. Some argue that energy constraints may limit token cost deflation, while others believe that pricing will improve with scale.
High adoption rates and productivity gains, with 11% of code shipped with zero human intervention at Uber.
Sticky token costs due to energy constraints, potentially limiting ROI and margin compression.