# Everyone's Talking About AI Capacity. Nobody's Talking About Why It's Breaking.

**Author:** WEKA

**Published:** August 17, 2026

![Black and white image of a data center with an open server rack and a long aisle of racks.](https://cdn.sanity.io/images/ult5g8gw/production/983d32745e9172893e27cb5fc13d42ba237bbc4b-1408x768.jpg)

## TL;DR

Expose why AI costs are spiking as agentic workloads collide with hard memory limits.

- Explain how subsidized AI pricing is ending and bills are surging.
- Show why GPU shortages mask a deeper, structural memory wall.
- Emphasize infrastructure and token efficiency as the new AI advantage.

The ride was subsidized. Now the meter is running.

Anthropic recently [moved its enterprise customers to usage-based pricing](https://gizmodo.com/anthropic-is-jacking-up-the-price-for-power-users-amid-complaints-its-model-is-getting-worse-2000746923), potentially doubling or even tripling bills for heavy users overnight. [Blackwell GPU prices jumped 48%](https://www.kucoin.com/news/flash/nvidia-gpu-rental-prices-surge-48-in-two-months-amid-ai-industry-s-worst-compute-shortage-in-five-years) in two months. [OpenAI’s internal thinking](https://the-decoder.com/openais-leaked-memo-says-new-spud-model-will-make-all-its-products-significantly-better/) is that whoever wins on compute wins the market.

The AI capacity crunch isn’t a forecast buried deep in the story anymore. It’s on the front page.

Eight months ago, WEKA said this was coming. Not as a vague warning, but a specific prediction: on stage, in the press, with a timestamp.

## The Uber model

Last fall during the [VentureBeat AI Impact Tour](/video/inside-the-ai-capacity-crunch-solving-latency-memory-limits-and-multi-agent-scaling-or), WEKA’s Val Bercovici made an analogy that holds up nicely: AI pricing in 2025 looked a lot like Uber in its early years, with heavily subsidized rides designed to hook users, build habits, and capture market share. The expectation was that once demand was locked in, the economics would eventually be forced to normalize with users paying real market rates to cover costs.

The conventional wisdom at the time was that AI would only get cheaper. [Token prices had dropped 280x over two years](https://medium.com/@horecny/the-ai-price-collapse-is-real-your-excuse-to-wait-is-not-cc575844497c). Everyone assumed the curve would just keep going. [Val pushed back on that](/resources/video/ai-inference-agent-swarms-and-token-economics-val-bercovici-at-venturebeat-ai-impact-tour/), saying organizations need to look at the net reality of tokenomics. Falling unit prices mask rising total costs. Agentic workloads — AI agents that plan, loop, call tools, spawn sub-agents — consume 10-100x more compute per user than conversational AI ever did.

[WEKA’s 2026 predictions](https://vmblog.com/prediction/weka-2026-predictions-subsidized-agent-era-ends-token-economics-mature-sovereign-infrastructure-emerges-and-enterprise-roi-becomes-measurable/) called it directly. The subsidized agent era would end. Token surge pricing would arrive. The free tier would evaporate. Enterprises would face a reckoning between what they budgeted for AI and what AI actually costs at production scale.

## The AI price reckoning is now

In April 2026, the surge pricing moment arrived, and it’s not isolated to one provider. This is an industry-wide recalibration.

Anthropic moved first. Enterprise customers on flat-rate seat plans — some paying up to $200 per user per month with bundled usage — are now being migrated to per-token billing at standard API rates. Analysts estimate this will double or even triple costs for heavy users. The company itself acknowledged the shift was tied directly to a deepening compute crunch.

The pressure is everywhere. OpenAI’s API usage went from roughly [6 billion tokens per minute](https://techcrunch.com/2025/10/06/sam-altman-says-chatgpt-has-hit-800m-weekly-active-users/) to [15 billion](https://openai.com/index/accelerating-the-next-phase-ai/) in only five months. [GitHub tightened Copilot limits](https://github.blog/news-insights/company-news/changes-to-github-copilot-individual-plans/). [OpenAI introduced caps](https://techcrunch.com/2026/04/09/chatgpt-pro-plan-100-month-codex/) on its most capable models and introduced a new $100/month tier to segment demand between its $20 Plus plan and $200 Pro ceiling.

Hardware costs are following the same trajectory. In addition to rising Blackwell GPU costs, [TrendForce estimates](https://www.trendforce.com/presscenter/news/20260202-12911.html) DRAM prices rose 90% to 95% quarter over quarter in Q1 2026, with NAND flash projected to rise 55% to 60%. [The memory supply shortage](/article/the-memory-shortage-exposes-broken-architecture-here-s-how-to-fix-it) is also expected to drag into 2027, adding to pricing pressure.

And here’s the thing that makes this structural rather than cyclical: Agents don’t just answer questions. They loop, plan, retry, reason, call external tools, and spawn sub-agents. The same user who consumed a modest number of tokens in a chatbot interaction consumes orders of magnitude more running an agentic workflow. As Val said at the VentureBeat event: “Can you afford all the tokens you want? You're vibe coding, generating cool apps, then you deploy in production and (suddenly) my bill is way too high.”

The market doesn't have fully mature rates for AI yet. But it’s starting to get them fast, and they’re higher than most enterprises planned for.

## Scale the memory wall

This isn’t a GPU shortage, but rather [a memory problem](/learn/ai-ml/ai-memory-wall/). And memory is governed by physics, not procurement budgets.

GPU memory grows significantly generation over generation, but it has struggled to keep pace with the explosive demands of new models. This gap isn’t closing; it’s widening. This is the memory wall, and it’s inside your infrastructure already.

A single 100,000-token context eats 40GB of HBM — that’s half the memory of an H100 gone, for one session. The [key-value cache (KV cache)](/learn/ai-ml/what-is-kv-cache/) fills fast. And when it fills, GPUs don’t wait politely; they recalculate work they’ve completed already, wasting compute cycles. Now run a swarm of agents doing this simultaneously, and your expensive compute cluster is essentially sprinting in place.

You can’t buy your way out of this. As WEKA CTO Shimon Ben-David put it during the [VentureBeat AI Impact tour](/video/the-agentic-ai-infrastructure-playbook): “Inferencing is not a GPU cycles challenge. It’s mostly a GPU memory problem. There are some problems that you cannot throw enough money at to solve.”

That’s not pessimism. It’s a constraint. And constraints are where architectural decisions either pay off or don’t. The teams hitting latency spikes, availability ceilings, and unpredictable costs right now aren’t underfunded. They’re running infrastructure that wasn’t designed for the workloads now living on it.

## The AI leaderboard built infrastructure for this moment

The competitive edge has moved. Model quality is table stakes. AI infrastructure economics is where the real advantage happens.

Token efficiency via tokenomics is the new competitive moat. It’s all about producing more tokens at a lower unit cost per token. The AI industry has made a transition: Software has gone from near-zero marginal cost to relatively high marginal cost. This changes everything about how you architect, deploy, and operate AI at scale.

Optimization is no longer a nice-to-have; it’s existential. The teams that treat AI infrastructure like an operating system, tuning every layer rather than just buying more capacity, are those that will come out of this repricing moment in a stronger position. KV cache acceleration, [memory-efficient inference architectures](/learn/ai-ml/how-ai-inference-works/), and purpose-built storage that keeps GPUs fed aren’t infrastructure luxuries. They’re the difference between AI at scale that either works to your advantage or eventually bankrupts you.

The subsidized rides are over. The question isn’t whether AI pricing will normalize. It’s whether your infrastructure can handle what comes next.

👉 Interested in additional insights on how to tackle the AI capacity crunch? Check out this [discussion on GPU capacity planning](/resources/video/gpu-capacity-planning-and-compute-market-dynamics/) with industry experts from Lambda, Silicon Data and Meta.