{
  "$type": "site.standard.document",
  "bskyPostRef": {
    "cid": "bafyreig3k55rkmc4twumfahlnyi5mnntpdno32ekvl22htj7nroi2sg7sm",
    "uri": "at://did:plc:mg5ozsljpp6t5b4lvwys4t72/app.bsky.feed.post/3mjztr2f7qvq2"
  },
  "coverImage": {
    "$type": "blob",
    "ref": {
      "$link": "bafkreihgqmfnxthtwu5d2jb6gatw6cf2e7xo54i532zvtzeqgw2yt7jbbi"
    },
    "mimeType": "image/gif",
    "size": 250078
  },
  "description": "OpenAI, Anthropic constraints and rising GPU prices signal demand is outpacing available compute\n",
  "path": "/ai-compute-shortage-challenges-bubble-narrative/",
  "publishedAt": "2026-04-21T20:40:42.000Z",
  "site": "https://broadbandbreakfast.com",
  "tags": [
    "Learn more about the Broadband Community...",
    "Start Your Broadband Journey Here"
  ],
  "textContent": "WASHINGTON, April 21, 2026 – Artificial intelligence companies are running out of compute capacity as demand surges, contradicting claims of an AI bubble, said**Vlad Galabov** , senior director at Omdia, Tuesday at Data Center World.\n\nMajor AI firms are already hitting limits. OpenAI scaled back its video-generation tool Sora earlier this year to redirect compute to core services as usage spiked.\n\nToken demand, the units of text processed by AI models, jumped from about 6 million tokens per minute in October 2025 to roughly 15 billion by March 2026.\n\nLearn more about the Broadband Community...\n\n\n                            Start Your Broadband Journey Here\n                        \n\nAt Anthropic, users of Claude Code reported hitting usage caps faster than expected in April. The company said addressing compute shortages tied to AI agents was a top priority.\n\nGoogle CEO**Sundar Pichai** warned as early as November 2025 that compute limits were already constraining growth.\n\nMeanwhile, xAI is scaling aggressively, expanding from more than 200,000 GPUs to 555,000 and targeting 1 million to train its latest models.\n\nGPU prices are rising despite increased supply. The cost of Nvidia compute time has risen about 48 percent in recent months, reflecting persistent shortages.\n\nThe “AI bubble” narrative confuses stock market behavior with real infrastructure demand, Galabov said.\n\n“There are two different concepts when I think about the aspect of an AI bubble. On one hand, you have the actual demand for chips. On the other hand, you have the stock market,” he said.\n\nArguments that more efficient models will reduce compute needs are not holding, Galabov said.\n\nAI workloads are becoming more complex, shifting to multi-step reasoning and agent-driven tasks that require more compute.\n\nAI agents are accelerating demand by triggering repeated model calls for a single task.\n\n“The demand really is there. The demand is not limited by the population of the world. It is limited by how many bots we decide to spin,” Galabov said.\n\nSupply remains limited by infrastructure bottlenecks, including power availability, grid constraints, shortages of memory and other components, and limited semiconductor manufacturing capacity.\n\n“The demand is not high because there is a bubble. The demand is high because we’re using it,” Galabov said.\n\nCompanies could generate more revenue if they had access to additional compute, he added.",
  "title": "AI Compute Shortage Challenges ‘Bubble’ Narrative",
  "updatedAt": "2026-05-22T21:49:26.392Z"
}