Divya van Mahajan

AI’s New Acceleration Phase

ai-daily-brief-podcast

AI’s New Acceleration Phase: Weekly Recap (May 22, 2026)

Overview

This episode of the AI Daily Brief — a daily podcast and video covering the most important developments in AI — presents a weekly recap framing the week of May 19–22, 2026 as a moment of surprising, multi-dimensional AI acceleration. The host argues that while individual stories might seem incremental in isolation, their collective weight signals a qualitative shift across business models, model capabilities, consumer services, policy, and infrastructure. The speaker is the host of the AI Daily Brief (name not stated in transcript).

Source video URL: Not provided.

Prerequisites

  • Basic familiarity with the major AI labs: OpenAI, Anthropic, Google DeepMind
  • Understanding of AI pricing models (flat-rate vs. usage/token-based billing)
  • Awareness of AI coding tools such as Claude Code (Anthropic), GitHub Copilot (Microsoft), and Codex (OpenAI)
  • General knowledge of AI hardware and compute infrastructure, including Nvidia and NeoClouds
  • Familiarity with terms like LLMs, agents, tokens, and recursive self-improvement (RSI)
  • Basic awareness of US AI policy debates and figures such as David Sacks

Main Points

1. Profitability Acceleration: Anthropic’s First Profitable Quarter

  • Anthropic projects its first-ever profitable quarter, marking the first time any major AI lab has reached this milestone.
  • Caveats include: the quarter is not yet complete (projection, not realized revenue); revenue recognition methodology counts top-line revenue before partner distributions; and the company has access to discounted compute from SpaceX for a limited period.
  • OpenAI also had a strong Q1, generating roughly $1 billion more in revenue than Anthropic, boosted significantly by the token-intensive Codex product.
  • Nvidia surpassed all analyst expectations, and with Jensen Huang’s cited forward demand pipeline of $1 trillion, analysts suggest Nvidia could be valued at $8–9 trillion — though investors remain uncertain how to price it at its current $5 trillion+ market cap.
  • The host interprets these results as resetting market expectations about the profitability of AI at scale, displacing the earlier “bubble” narrative.

2. End of the Subsidy Era: Shift to Usage-Based Pricing

  • The host frames the current moment as the transition from a “subsidy era” (flat-rate plans that subsidize heavy users) to a “trade-off era” (usage-based billing driven by token-hungry agentic workloads).
  • Google I.O. announced a price reduction on its Gemini Ultra plan from $250 to $200/month, but this came alongside a shift to usage-based billing for token-intensive tasks — structurally similar to Anthropic’s earlier pricing changes.
  • Microsoft canceled its enterprise Claude Code licenses, citing cost considerations alongside competitive incentives to improve GitHub Copilot.
  • As one market commentator put it: “Token-based pricing is forcing every enterprise customer to confront the actual cost of running these models at scale, and the number turns out to be far higher than the flat-rate experiments suggest.”
  • Anthropic introduced a /usage command in Claude Code so users can identify which agents, MCPs, or plugins are consuming the most tokens.

3. Compute Supply Acceleration: SpaceX as an AI Compute Provider

  • Elon Musk publicly announced that SpaceX is offering AI compute as a service at significant scale, citing the expanded Anthropic partnership as a demonstration.
  • Anthropic’s Chief Compute Officer Tom Brown confirmed the company is scaling up on both Colossus 1 and the new Colossus 2 data center.
  • The host argues this positions SpaceX as an emerging alternative NeoCloud, with a unique long-term possibility of expanding compute capacity via orbital data centers — something no other NeoCloud can offer.
  • This development is expected to increase investor interest in the SpaceX IPO, shifting its narrative from an “Elon Musk vehicle” to a strategic AI infrastructure company.
  • Separately, Cursor released Composer 2.5, a coding model benchmarked as comparable to Opus 4.7 and GPT-5.5 but at 10–60x lower cost, representing a software-side response to the compute cost problem.
  • The Gemini app reached 900 million monthly active users, nearly closing the gap with ChatGPT.
  • Monthly tokens processed by Google grew 700% year-over-year, from 480 trillion to 3.2 quadrillion.
  • Google announced persistent AI agents integrated directly into Search: users can create ongoing information agents that monitor the web, news, finance, sports, and social media and deliver synthesized updates when relevant information appears.
  • The host identifies this as potentially one of the most significant Google I.O. announcements because it transforms Search from a one-time query tool into a persistent information-gathering service.
  • Google also announced Docs Live, allowing users to create and edit documents via voice plus AI, signaling an accelerating shift toward voice-first, live interaction patterns.

5. Model Capability Acceleration: AI Solves an 80-Year-Old Mathematics Problem

  • OpenAI announced that an internal model solved a problem posed by mathematician Paul Erdős in 1946: given n points on a plane, how many pairs can be exactly one unit apart?
  • The prevailing belief had been that a square grid arrangement was optimal. The AI model disproved this conjecture using multidimensional mathematics projected into 2D space.
  • Fields Medalist Tim Gowers described it as “the first really clear example of AI solving not just an unsolved math problem, but a really well-known unsolved math problem.”
  • Notably, the model used was a general-purpose LLM with no special mathematical training, and the solution required no unusual prompting — just a clear statement of the problem.
  • OpenAI researcher Alexander Wei framed it as a leading indicator: “Soon, perhaps sooner than we all think, AI will begin autonomously producing landmark results in CS, physics, econ, bio, etc.”
  • Energy cost estimate (Ethan Mollick): solving the problem required the equivalent of 2–20 miles of EV driving in electricity and less than 3 almonds’ worth of water.

6. Research Acceleration: Andrej Karpathy Joins Anthropic to Work on RSI

  • Former OpenAI co-founder Andrej Karpathy announced he is joining Anthropic, describing the next few years at the LLM frontier as “especially formative.”
  • His focus will be recursive self-improvement (RSI) — specifically, using Claude to accelerate pre-training research itself.
  • Anthropic’s Nicholas Joseph stated Karpathy will build a team focused on this goal.
  • The host frames this as a significant signal: a researcher already worth billions choosing to return to active work at this specific moment and on this specific problem.

7. Data Center Politics: Opposition and Counter-Narrative Acceleration

  • Opposition to data centers is accelerating as a political issue in communities across the US.
  • Simultaneously, a counter-narrative is gaining momentum, pushing back on the most-cited concerns:
    • Water use: annual data center water consumption is less than one-fifth that of golf courses, one-tenth that of almond farming, and one-twentieth that of lawns.
    • Job creation: local reporting (e.g., a CBS affiliate in Richmond, Virginia) highlights significant employment benefits for tradespeople like electricians building AI infrastructure.
  • The host expresses support for community debate grounded in factual evidence rather than general fear of AI or big tech.

8. Policy Acceleration and Setback: The Scuttled AI Executive Order

  • California Governor Gavin Newsom signed an executive order directing state agencies to study and prepare for AI-related labor disruption, tasking them with gathering data, identifying early warning signs, and exploring reforms to severance and unemployment insurance.
  • The order is exploratory in nature; critics noted the practical difficulty of attributing layoffs specifically to AI through existing state unemployment systems.
  • At the federal level, the Trump administration appeared set to sign a sweeping AI executive order that would have required AI companies to submit models to the government 90 days before public release (industry pushed for 14 days) and established public-private security hardening protocols.
  • The order was canceled hours before the signing ceremony. Former AI czar David Sacks reportedly intervened personally, telling Trump that companies were already cooperating voluntarily and that mandatory pre-release review would slow innovation and hurt the US in its competition with China.
  • Trump stated: “I didn’t like certain aspects of it, so I postponed it. I think it gets in the way.” He repeatedly framed the issue around not losing ground to China.

Key Concepts

  • Subsidy era: The prior period in AI consumer pricing in which flat-rate subscription plans effectively subsidized the heaviest users, made viable by relatively low per-query costs.
  • Trade-off era: The emerging pricing paradigm in which token-hungry agentic workloads make flat-rate subsidization economically unviable, forcing a shift to usage-based billing.
  • Token-hungry agents: AI agents that consume very large numbers of tokens to complete tasks, dramatically raising the cost of flat-rate service plans.
  • Recursive self-improvement (RSI): A research paradigm in which AI systems are used to accelerate the research and development of future AI systems, potentially creating a self-reinforcing improvement cycle.
  • NeoCloud: A category of cloud compute provider focused specifically on AI workloads, distinct from hyperscalers like AWS or Azure (examples: CoreWeave, Lambda Labs; SpaceX is framed as an emerging entrant).
  • Orbital data centers: A proposed future infrastructure concept in which data centers are deployed in space, potentially enabling compute capacity expansion beyond terrestrial constraints.
  • Persistent search agents: AI agents embedded in search engines that continue monitoring the web on a user’s behalf over time and deliver synthesized updates, as opposed to one-time query responses.
  • Erdős unit distance problem: A combinatorics/geometry problem posed in 1946 asking for the maximum number of unit-distance pairs among n points in a plane; the prevailing square-grid conjecture was disproved by an OpenAI model this week.
  • Claude Code / Cloud Code: Anthropic’s coding-focused AI tool, referenced as a token-intensive developer product with ongoing enterprise adoption questions.
  • Composer 2.5: Cursor’s coding model, benchmarked as performing comparably to frontier models at 10–60x lower cost.

Summary

The host argues that the week of May 19–22, 2026 represents not a collection of discrete news items but a convergent acceleration across every major dimension of AI development: AI labs are becoming profitable for the first time, the business model of AI is structurally shifting from subsidized flat-rate to usage-based pricing, new compute infrastructure from unexpected players like SpaceX is coming online, consumer AI services are reaching mass scale and gaining new agentic capabilities, a general-purpose language model has solved a famous 80-year-old mathematics problem with minimal effort, and one of the field’s most prominent researchers has returned specifically to pursue recursive self-improvement. Counterposed against this technical and commercial acceleration is a policy environment in flux — a California labor-disruption order of uncertain practical effect, and a federal AI executive order scuttled at the last minute by deregulatory pressure. The host concludes that the cumulative feeling of this week — captured in Demis Hassabis’s framing of humanity standing “in the foothills of the singularity” — is one of standing at the threshold of a genuinely new phase in AI development, and that the challenge now is ensuring its benefits are broadly realized.