Divya van Mahajan

The Annual AI Slowdown Panic is Here

ai-daily-brief-podcast

The Annual AI Slowdown Panic Is Here

Overview

This episode of the AI Daily Brief podcast (published May 27, 2026) argues that the recurring phenomenon of summer AI slowdown narratives has arrived early in 2026. The host examines three major topics: a new agentic coding benchmark called DeepSWE, a shift in how AI industry leaders discuss job displacement, and the funding surge in AI inference infrastructure. The central thesis is that the current “token crunch” and signs of slowed growth in certain metrics are being misread as evidence of a bursting AI bubble, when in fact they reflect a healthy market transition from subsidized experimentation to sustainable, usage-based economics.

Source video URL: not available

Prerequisites

  • Basic familiarity with large language models (LLMs) and generative AI products (e.g., ChatGPT, Claude, Gemini)
  • Understanding of AI benchmarks and why they matter for model evaluation
  • General awareness of agentic AI coding tools (e.g., Claude Code, OpenAI Codex, Cursor)
  • Familiarity with concepts like inference vs. training, tokenization, and API-based AI consumption
  • Basic economics concepts: supply and demand, market pricing, bubble dynamics

Main Points

1. DeepSWE: A New Coding Benchmark Designed to Resist Gaming

  • Most existing benchmarks (e.g., SWE-Bench family) scrape real GitHub issues and PRs, creating two problems: memorization (models may have seen solutions in training data) and triviality (tasks tend to be small and simple).
  • DeepSWE, created by DataCurve, builds tasks from scratch with short, natural prompts that require significantly more code to solve, reflecting realistic long-horizon engineering work.
  • Initial results show GPT-5.5 leading at 70%, GPT-5.4 at 56%, and Claude Opus 4.7 at 54%. Chinese models scored far lower (Kimi K2.6 at 24%, DeepSeek V4 at 8%), revealing a divergence hidden by other benchmarks.
  • GPT-5.5 also led on cost and token efficiency — using roughly half the tokens of Opus 4.7, completing tasks in less than half the time, at about one-third the cost.
  • A key qualitative finding: top models (GPT-5.4 and Opus 4.7) self-verified their work by writing their own tests more than 80% of the time; weaker models rarely did this. A distinct failure pattern for Claude was identified: missing requirements in multi-part prompts (e.g., implementing sync but not async).
  • DataCurve withholds solutions from GitHub to prevent contamination of future training data, a methodological safeguard praised by the community.

2. AI Leaders Revising the “Jobs Apocalypse” Narrative

  • OpenAI CEO Sam Altman stated he no longer believes in the kind of jobs apocalypse that some AI companies have predicted, acknowledging that the impact on entry-level white-collar jobs has been less than he anticipated.
  • Altman attributed this to underestimating the irreplaceable human dimension of employment: “we really do care about our interactions with people.”
  • Goldman Sachs CEO David Solomon published a New York Times op-ed arguing the AI job apocalypse is overblown, citing Goldman economists who estimate ~25% of work hours could be automated over the next decade.
  • Solomon’s argument follows historical precedent: markets typically use productivity tools to deliver better products at the same price, not the same product at lower cost — generating new jobs and a productivity boom rather than mass displacement.
  • The host notes that economists have long distinguished between task automation and job automation, and that organizational friction creates a natural speed limit to AI-driven displacement.

3. The Inference Layer Attracts Major Funding

  • Base10, a vertically integrated platform for fine-tuning and deploying open-source models (acting as a value-added reseller for cloud GPU capacity), is closing in on a $1 billion fundraising round valuing it at $11 billion — more than double its valuation from just three months prior. Annualized revenue tripled from $200M to $600M in Q1 alone, representing a 20x increase from March of the previous year.
  • OpenRouter, a token-routing service enabling developers to access multiple AI models through a single API, raised a $113M Series B led by Capital G (Alphabet’s investment arm), valuing it at $1.3 billion. The company now serves 100 trillion tokens per month — a 5x increase in six months.
  • The host quotes investor Dylan Brislott: the key insight is that training runs are amortized, but serving runs repeat every time a user opens an app — the marginal dollar in 2026 goes to inference, not training.

4. The Pattern of Annual AI Slowdown Panics

  • The host identifies a recurring seasonal pattern of AI pessimism:
    • Summer 2023: ChatGPT’s first traffic decline attributed to novelty wearing off; later explained by students going home for summer.
    • Summer 2024: Concerns about a “pre-training data wall” — models allegedly running out of training data and facing an improvement ceiling.
    • Summer 2025: A widely cited MIT “study” claimed 95% of generative AI projects fail; GPT-5 disappointment fueled a “bubble” narrative; resolved when Claude Code, Opus 4.5, and GPT-5.3/5.4 demonstrated major capability leaps.
  • Each panic eventually resolved as new capabilities or business models emerged.

5. The Token Crunch and the End of the Subsidy Era

  • The first half of 2026 was characterized by explosive agentic AI adoption: companies like Uber burned through annual token budgets in four months; OpenAI reached a ~$30B revenue run rate; Anthropic reached ~$45B.
  • This growth strained infrastructure: tokens became scarce and expensive, forcing a transition from seat-based subscription pricing to usage-based (pay-per-token) pricing.
  • Some prosumer users consuming $5,000–$10,000 of tokens on $200/month plans are being directly affected. The host describes this as the end of “the AI subsidy era.”
  • The U.S. government reportedly blocked Anthropic from expanding access to its most powerful model (“Mythos”), partly to secure priority access to its token capacity.

6. Why the Current Bubble Narrative Is Likely Wrong

  • Uber’s COO stated that high token spending had not translated into more useful consumer features — which critics used to build a narrative of impending collapse.
  • The host and cited commentators counter with several data points:
    • GPU rental prices are up 2x in four months, indicating demand is still significantly outrunning supply — the opposite of a collapsing market.
    • Epoch AI estimates global inference supply is tripling annually, while demand is growing ~10x annually — a structural shortage, not a glut.
    • A viral chart showing a plateau in VS Code AI plugin installs is misleading: it does not count users of terminal-based tools (CLI) like Codex. NPM installs of Codex grew from ~100,000/day in January to over 1.5–1.8 million/day — suggesting platform shift, not demand collapse.
  • Professor Ethan Mollick summarized the dynamic: rising demand leads to higher costs, which modulates demand — a normal market equilibrium, not a crash.

7. What Is Genuinely New and Worth Discussing

  • The end of free experimentation creates real risks: reduced ability for non-technical users to explore novel agentic workflows, and potential AI inequality where only well-resourced organizations access the most capable models.
  • However, resource constraints also slow the pace of AI-driven job displacement, buying time for organizational and social adaptation — which the host views as a market-driven, healthier alternative to a top-down mandated slowdown.
  • A new concept emerging in practitioner circles: “agent debt” — the accumulation of messy, unconsolidated agent workflows that create compounding technical problems over time, analogous to technical debt in software development.
  • Both OpenAI and Anthropic have launched consulting ventures to help organizations deploy AI more thoughtfully.

Key Concepts

  • DeepSWE: A new agentic coding benchmark by DataCurve built from scratch to test long-horizon, real-world software engineering tasks, designed to resist memorization and gaming.
  • Self-verification: A behavior in top-performing coding models where the model writes its own tests to validate its output, identified as a key differentiator between leading and lagging models.
  • Token crunch: The current state of AI infrastructure in which demand for AI inference tokens substantially outpaces supply, leading to rationing, price increases, and business model changes.
  • Subsidy era: The period (roughly early–mid 2026) when AI labs effectively sold access to compute below cost to drive adoption, now ending as usage-based pricing takes hold.
  • Inference layer: The infrastructure layer responsible for running trained models to generate outputs for end users — now the primary focus of AI investment in 2026.
  • Base10: A neo-cloud company providing vertically integrated fine-tuning and deployment of open-source models as a value-added reseller of GPU capacity.
  • OpenRouter: A token-routing service allowing developers to access multiple AI models through a unified API, optimizing for cost, performance, or redundancy.
  • Agent debt: Technical debt specific to agentic AI systems — accumulated from rapidly assembled, uncleaned agent workflows that degrade in reliability and coherence over time.
  • Punctuated equilibrium: A pattern of growth characterized by periods of stability punctuated by rapid spikes, used here to contextualize apparent plateaus in AI tool adoption metrics.
  • Task automation vs. job automation: The distinction that automating individual tasks within a job does not necessarily eliminate the job itself, due to the bundled human and relational dimensions of most roles.

Summary

The host argues that the AI industry in mid-2026 is experiencing another iteration of its annual summer slowdown panic — this time triggered by the end of the subsidized token era, a token crunch forcing companies from flat-rate to usage-based pricing, and isolated reports of AI spending not yielding clear returns (e.g., Uber’s COO comments). Rather than signaling a bursting bubble, the host contends these dynamics reflect a normal and healthy market correction: GPU prices doubling and inference demand growing at roughly 10x annually demonstrate that AI compute remains severely supply-constrained, which is structurally incompatible with a collapsing demand story. Apparent slowdowns in metrics like VS Code plugin installs are better explained by platform migration to CLI-based tools. The host acknowledges genuine costs to the transition — reduced experimentation capacity and risk of AI inequality — while arguing that resource-constrained markets forcing sustainable pricing are far less bubble-prone than subsidized ones. The episode closes with a call to treat this period constructively, addressing emerging challenges like agent debt and organizational AI adoption, rather than cycling through another round of premature bubble declarations.