Beyond the Subsidy

May 8, 2026 · Technology

There's a narrative forming around recent changes in AI pricing, that the subsidization era is ending, that companies are tightening the screws, and that the free ride is coming to an end. The framing tends to be financial, where companies are losing too much money and need to start charging users more.

That framing gets the symptom right but the cause wrong. The AI economy isn't straining because these companies are running out of money. It's straining because they're running out of compute.

How the Cracks First Appeared

The subsidization problems started earlier than most people realize. The first major break came with Cursor in mid-2025.

Cursor had been pricing by the number of messages, where you pay a monthly fee for a fixed count. The problem is that messages aren't uniform, ranging from a few cents for a simple query to many dollars for a long agent run. A developer running a complex task can spend as much on a single request as a casual user spends in an entire month.

Cursor doesn't own its own compute. It pays Anthropic and OpenAI for every inference call, so when some users were extracting $10 of compute with their 200 monthly messages while others were pulling $1,000 or more, Cursor couldn't absorb that spread. They switched to usage-based pricing, where what you pay reflects what it actually costs to serve you. GitHub Copilot still prices by message count, and they're now going through the same reckoning Cursor faced a year earlier.

Anthropic's Compute Steering Experiments

A few months before the more recent pricing drama, Anthropic ran a quiet experiment. They announced they'd be doubling usage limits outside of peak hours for a two-week period, with the goal of pushing heavy users toward off-peak times and freeing up GPU capacity during the business day.

The reasoning is straightforward if you think in terms of fixed resources. Anthropic has a finite number of GPUs, and during peak hours most of them are saturated, making the marginal cost of one more request at 2pm very different from the cost of one at 2am. The experiment didn't achieve what they hoped. A week and a half later, Anthropic announced they'd be adjusting 5-hour session limits during peak hours, consuming usage faster in the windows when demand was highest. They'd tried carrots, then moved to sticks.

This context matters for understanding the later change that got more attention, the move to restrict Claude Code from the $20/month tier. The common framing was that Anthropic wanted to push users up to the $100 plan, but that's the wrong read.

Anthropic doesn't care much about an extra $80 from $20 subscribers, because their subscription revenue is largely a marketing play, not a profit center. Enterprise customers are where the real money is, and the $20 plan change was the same move as the off-peak experiment, just more aggressive, throttling the consumer tier to claw back compute capacity for the customers actually paying commercial rates.

The Subsidization Math

To understand how unusual the current situation is, consider what heavy users on consumer subscriptions are actually extracting. A $200/month Claude subscription can generate thousands of dollars in raw compute value. One internal measure found that users can pull $5,000 or more in inference from that plan, because Anthropic is subsidizing that aggressively.

The raw infrastructure cost to run model inference is estimated at roughly 15-20% of the API price, so if you're paying $15 via the API, that costs Anthropic around $1.50-$3.00 in electricity and compute. That math works. On the subscription model, though, when some users are burning through compute at a 20x or 25x ratio relative to what they're paying, it breaks down completely, with some users costing Anthropic money on pure electricity alone before factoring in hardware, training, or any overhead.

As a rough illustration, a single consumer-grade GPU running at full load 24/7 raises an electric bill by roughly $500-$1,000 a month depending on location. Data center electricity is cheaper, but not by as much as people assume, and the subsidization these companies are absorbing is real money.

This was never meant to be sustainable indefinitely. It's a deliberate marketing investment designed to maximize developer adoption and mindshare, but as compute gets tighter, that investment becomes harder to maintain.

Pre-training vs. Post-training

One nuance that often gets lost involves the difference between pre-training and post-training costs. A conversation with Anthropic CEO Dario Amodei laid out a useful framework for thinking about model economics that most coverage missed.

Pre-training is the expensive phase, where terabytes of data get compressed into a model's parameters to produce a fundamentally new model. This is where the hundreds of millions or billions of dollars go.

Post-training is different. It's the process of refining how a model behaves, using reinforcement learning strategies like RLHF and RLVR to make models better at specific tasks like long agent runs, tool use, and code generation. Post-training has proven far more impactful than anyone expected. Models like Cursor's Composer 2 aren't new pre-trains; they're fine-tuned versions of existing open-weight models, and the compute required is generally much cheaper than a full pre-training run.

This matters when evaluating claims about certain model releases losing money. When a model update feels fundamentally different, it's usually new pre-training. When it arrives faster, runs cheaper, and is slightly smarter in focused areas, it's usually post-training on top of an existing base. A model that felt like a major leap probably cost near a billion dollars; the follow-on updates came in considerably cheaper.

Microsoft's Clearest Signal

The GitHub Copilot situation is the most illustrative example of what's actually happening, and Microsoft's decision to pause new signups is the single detail that makes it clearest.

Copilot had been pricing by number of requests, which has the same fundamental problem as Cursor's early model, where different models have wildly different costs, and different requests within the same model have wildly different costs. A long agent task can cost more than thousands of simple chat messages. The multipliers Copilot now shows for different models, GPT-5.4 at 1x, GPT-5.5 at 7.5x, and Claude Opus 4.7 at 15x, aren't based on the actual API cost differential. GPT-5.5, despite being listed at 7.5x the message cost, is actually cheaper to run than GPT-5.4 in terms of effective cost per task, because it uses far fewer tokens.

These numbers aren't tracking electricity costs; they're tracking availability. How many GPUs are provisioned for each model, and how much of that capacity does Microsoft need to reserve for enterprise customers paying full commercial rates? The message to consumers isn't "we need more money from you." It's "we need the GPUs that Copilot is consuming available for the Azure customers paying real rates."

The signup pause makes this explicit. You don't pause signups because you want to make more money. You pause signups because you don't have capacity.

OpenAI did the same thing years ago after a product launch surge. They paused ChatGPT Plus signups because they'd exceeded their available compute, solved it by buying every GPU they could find, and once the hardware arrived, the restrictions lifted. Anthropic and Microsoft are in the same position now, and the restrictions will ease when more compute comes online.

Google is the most commonly overlooked case and arguably the most extreme example of the subsidization crunch.

Google doesn't get the same attention as Anthropic or OpenAI in these conversations, partly because their models have been weaker and partly because developers in the AI tooling space tend to ignore products they don't personally use. Google has been doing more free inference than anyone else, by a significant margin. AI Overviews in Google Search are free compute served to billions of queries, including users who aren't signed in. Google's Antigravity IDE included access to Claude's best models for a period, subsidized so aggressively that people were subscribing specifically to get cheap Claude inference. That offer attracted enough heavy users that it exploded, and Google was among the first companies to crack down, banning plugins that automated usage, restricting integrations with coding tools, and cutting off anything that turned their subscription into an inference pipe.

The reason Google gets overlooked is that they went through the crunch first and already tightened before most people were paying attention. Because their developer tooling isn't widely adopted, the restrictions didn't generate the same complaints that Anthropic's changes did, but the pattern is identical. Over-subsidize to gain adoption, hit a compute ceiling, restrict.

The Cost of Intelligence Is Falling

The claim that compute per dollar is declining isn't wrong, but it needs precision. Frontier model prices at the bleeding edge are going up, but at any fixed level of capability, prices are dropping fast.

GPT-5.4 represented a certain benchmark of capability at a certain price. About a month and a half later, GPT-5.5 medium matched that capability at less than half the cost, and GPT-5.5 low runs at roughly a sixth the price of GPT-5.4 on equivalent tasks while still scoring competitively against models that cost much more just months ago.

The efficiency gains are visible in token counts, not just prices. GPT-5.5 medium solves the same benchmark tasks using about 22 million tokens where Sonnet 4.6 uses 200 million, and GPT-5.5 low achieves the same score as Deepseek V4 using only 7 million tokens compared to Deepseek's position at the top of the chart. That efficiency gap represents real cost reduction for anyone measuring what it actually costs to complete work, not just what it costs per token.

This also shows why the Copilot multipliers make no sense as cost signals. GPT-5.5 is listed at 7.5x the message cost of GPT-5.4, despite being significantly cheaper on real tasks. The pricing isn't tracking electricity, it's tracking compute availability and who Microsoft needs to protect that capacity for.

Enterprise vs. Consumer

When companies like Uber report spending their entire annual AI budget in four months, they're not using personal subscriptions. They're hitting the API directly, paying commercial rates that are much higher than what benchmarks report, because enterprise contracts include SLAs, support, data handling agreements, and zero data retention.

Consumer subscriptions have explicit clauses prohibiting commercial use at scale, and most companies won't permit it anyway, since consumer plans don't offer HIPAA compliance, proper data isolation, or the BAA agreements that regulated industries require. There are engineers at companies spending more on API inference than they earn in salary, because their companies are using the API directly at full commercial rates, not $200 subscriptions.

The consumer tier exists as a marketing channel; the enterprise tier is the actual business. When Anthropic restricts consumer access, they're not pushing individuals up to higher subscription tiers. They're protecting the compute capacity that enterprise customers pay for at rates that actually make commercial sense.

What's Actually Happening

The limited resource isn't money. It's GPUs.

Misreading compute scarcity as revenue-seeking leads down the wrong path. You don't pause signups to make more money. You don't give users $5,000 of inference for $200 if you're trying to maximize revenue. You don't give away free AI overviews to billions of signed-out users if you're running a tight profit operation.

These are the moves of companies fighting over a constrained physical resource and trying to allocate it toward the customers who matter most to their actual business. The restrictions affecting consumers aren't the end of generous AI access, just the market finding the edges of the hardware supply.

The subsidized compute era isn't over; it's contracting to fit within the supply that actually exists, and it will expand again as more compute comes online. Understanding that difference changes how you interpret every pricing change that comes next.