GitHub Copilot abandons fixed-seat subscriptions for usage-based billing + an analyst burns $6,000/day in Claude tokens to independently replicate a 100-person team's decade of work + Uber's CTO exhausts the annual AI budget ahead of schedule—three seemingly unrelated events that share one structural shift: the SaaS "seat subscription" model is being displaced by "compute billing", and AI capabilities are being "grid-ified"—transforming from software licenses into metered cognitive public infrastructure.
AI is not a cost-reduction tool; it is a lever that rewrites organizational architecture. In a world where one $6,000/day analyst replaces 100 analysts: management layers will flatten, the leverage of a few high-value employees will amplify a thousandfold, and budgets will shift from "headcount costs" to "compute costs"—HR departments shrink, FinOps departments grow.
The initial narrative of "burning through the budget prematurely" has, five months later, been flipped by Uber itself into a governance case study of "34% unit cost reduction" (Axios); Microsoft and OpenAI's revenue sharing has also crystallized into a contractual clause with a clear cap (cumulative $38 billion through 2030, CNBC). Compute billing did not end in runaway costs—it entered a "governable" next phase. See the update section at the end for details.
GitHub Copilot's shift to usage-based billing sparked a lively discussion with 486 points on HN. On the surface, this is just one company's pricing strategy adjustment, but it marks a larger industry turning point. The seat subscription model—$19/person/month, regardless of usage—is the core commercial logic of the SaaS era, predicated on software being a fixed-cost infrastructure where price should be tied to user count rather than usage. But AI tools break this premise: the usage variance among different developers on Copilot can exceed 100×, with heavy users causing vendor losses while light users contribute disproportionate profits.
Usage-based billing isn't just fairer to vendors; it reveals a deeper reality: AI capabilities are being "grid-ified"—transforming from software licenses into metered cognitive infrastructure. Remunerationlabs' analysis uses a precise metaphor: AI tokens are undergoing a "grid-ification" evolution—from the customized, isolated single-model era toward standardized, metered cognitive public infrastructure. "Tokenmaxxing" is becoming the productivity metric for elite engineering teams, and CFOs are replacing SaaS license budgets with token budgets.
Update (2026-09): This transition has moved from "announcement" to "implementation." GitHub's official changelog confirms that all Copilot plans switched to "GitHub AI Credits" token-consumption billing on June 1, 2026; base subscription prices remain unchanged (Pro $10, Business $19, Enterprise $39 per user/month), but these numbers now represent "included allowances" rather than usage caps—heavy users need to purchase additional usage, and annual subscribers who don't switch to monthly billing upon renewal are downgraded to the free tier. This confirms the judgment in this section: seat prices are the surface; token consumption is the new pricing anchor.
The most thought-provoking case from that week in April 2026 comes from The Neuron's report: SemiAnalysis analyst Dylan Patel recounted the story of "Jeremy"—one person, $6,000/day in Claude token costs, independently rebuilding in 3 weeks what a 100-person data services team took a decade to produce. That same week, Meta laid off 8,000 people, many of them middle managers and non-Agentic project researchers.
These two events together form the core proposition about labor value in the AI era: the winners in the AI era are not "people who master AI tools," but "people with deep enough domain judgment to convert AI compute into alpha value". The difference between Jeremy and his laid-off peers is not tool access—they likely all used Claude—but irreplaceable domain cognitive density. This proposition is corroborated by Epoch AI's data: Claude users skew heavily toward high-income groups, while Meta AI users skew toward low-income groups—those who can convert compute into alpha value are already concentrated in high-income, high-cognitive-density roles.
a16z's analysis further confirms: the most mature enterprise AI deployments concentrate on customer service, code generation, and internal knowledge retrieval—scenarios that share the characteristic of "quantifiable output, correctable errors." Truly high-value decision-making work—strategic judgment, creative breakthroughs, domain insights—remains the territory of the Jeremies, not a target for AI replacement. The question enterprise managers need to ask therefore shifts from "which roles will AI replace" to "how many Jeremies do we have".
When AI tokens become the new kilowatt-hour, the essence of enterprise competitiveness is cognitive efficiency. — Core judgment of this article
GAI Insights' in-depth report reveals OpenAI's commercial structure: establishing a $10 billion SPV called "DeployCo," partnering with PE giants like TPG, Bain, and Advent, promising partners a minimum 17% annual return over 5 years, targeting mid-market portfolio companies of PE funds. Simultaneously, Google Cloud is paying McKinsey, Accenture, and Deloitte $750 million through a parallel fund to promote AI—AI vendors are paying distributors, a logic that completely upends the traditional software sales model, analogous to Apple subsidizing AT&T to push the iPhone, or pharma companies paying high rebates to PBMs (see companion reading "OpenAI Is Paying Wall Street to Sell Its Software").
Tom Tunguz's analysis of Anthropic's strategy reveals the other side: Anthropic is executing the classic "commoditize the complement" strategy—offering certain SaaS capabilities for free or at very low cost, shifting enterprise budgets from SaaS licenses to inference compute. The two streams converge on the same direction: the enterprise software market in the AI era is being squeezed from both ends—AI vendors directly replacing SaaS functions at the top, and vendors paying channel partners to penetrate downward at the bottom. The middle layer of traditional SaaS companies is being evaporated, closely aligning with a16z's judgment.
On April 27, 2026, Microsoft stopped its revenue-sharing agreement with OpenAI. On the surface, this is a contract renegotiation; in substance, it is a rebalancing of power: when OpenAI's capabilities and brand are strong enough to no longer need Microsoft's channel endorsement, the revenue-sharing agreement becomes a bargaining chip between two equal parties rather than a compensation mechanism. OpenAI traded greater independence for the concession of revenue sharing, while Microsoft retained its equity stake and continues to receive a share of compute consumption on Azure.
Update (2026-09): The specific terms of this "partnership upgrade" have been finalized. According to CNBC, the revised agreement eliminates Microsoft's exclusive access to OpenAI models and IP (OpenAI can now license models to Microsoft competitors like Google and Amazon), and Microsoft no longer pays revenue sharing to OpenAI; however, OpenAI must still pay Microsoft a 20% revenue share, with the total cap locked at $38 billion, covering through 2030—a significant reduction from the previously expected ~$135 billion. In practical terms, OpenAI is projected to pay Microsoft about $6 billion in 2026, roughly one-fifth of its projected $30 billion full-year revenue. This "power rebalancing" is ultimately a quantifiable concession, not a one-time decoupling.
The deeper implication of this structure is: the relationship between AI vendors and cloud providers is shifting from "dependency-sustenance" to "mutual lock-in". Anthropic's $40 billion investment from Google is essentially the same logic: Google's money locks in Anthropic's compute spending, and Anthropic's model capabilities feed back into Google Cloud's competitiveness—"compute circular reflux" becomes a new type of strategic binding tool, more durable than contracts and more flexible than compliance requirements. For enterprise users, this means choosing an AI service provider is effectively choosing an entire infrastructure ecosystem, and the "switching cost" curve is becoming increasingly steep.
First, from seat budgets to compute budgets—finance departments need to manage AI token consumption the way they manage electricity consumption, establishing usage baselines, setting alert thresholds, and allocating budgets by business scenario. Second, identify the Jeremies in your organization—high AI-leverage talent isn't necessarily AI experts, but people who combine domain depth with AI tool fluency, whose output may be 10-100× that of ordinary employees. Third, understand the strategy behind "free"—when AI vendors offer free or low-cost SaaS replacement features, today's free functionality may be tomorrow's switching barrier. Fourth, establish a quantifiable AI ROI framework—the lesson from Uber's CTO shows that AI investment without ROI constraints is dangerous; successful enterprises are already versioning, auditing, and making AI workflows reusable.
The final form of this business model reconstruction is not yet settled, but one thing is clear: in the AI era, the core of competitive advantage is no longer "who uses the best model," but "who most efficiently converts compute into business value." GitHub Copilot is the first mainstream product to cross this line—Uber's CTO burning through the budget early + the $6K/day analyst are not anomalies; they are the norm of this new world (update: five months later, Uber has turned "budget blowout" into a governance exemplar of "34% unit cost reduction," see next section). In the next 12 months, Cursor / Replit / Notion / other SaaS will cross this line one after another.
The "Uber CTO exhausted the annual AI budget ahead of schedule" originally cited in this article comes from The Information's April 2026 interview with Uber CTO Praveen Neppalli Naga, and was referenced by Axios' report the same month. This story continued to evolve: in June 2026, Uber began setting usage caps on interactive sessions for tools like Claude Code (session cap of 400K tokens, even though the model can handle 1M tokens), routing tasks to different models based on cost and capability requirements.
According to Axios' exclusive report on August 27, 2026, the results were tangible: Uber's weekly AI agent request volume grew 9.4× compared to February, AI agents currently contribute over 70% of code commits, and engineers run over 30,000 AI agent tasks daily; yet total AI spending has remained stable since April—under the same model, cost per 1,000 requests dropped nearly 34% from the April peak, and cost per session dropped 52% from the June high. As Uber's CTO himself put it: "We're moving out of the so-called tokenmaxxing era."
This reversal deserves more attention from enterprise decision-makers than the initial "budget blowout": compute billing is not an inherently runaway black hole, but a new curve that requires active governance—model routing, session caps, tiering by task complexity are essentially the same moves FinOps applied to cloud computing, now reapplied to token consumption. This fully aligns with the framework in Section 5: unconstrained AI investment is dangerous; constrained AI investment is sustainable leverage.