Loading
Loading
The edge sits where the binding constraint just moved. Today that is power and the interconnect inside compute — raw model capability is no longer the scarce thing.
Updated May 24, 2026 · ranked by where the constraint sits
The scarce input for net-new capacity is energized land. Multi-year interconnection queues, transformer lead times, and large-load tariff fights set a data center’s schedule more than GPU allocation does. Hyperscalers are paying up for behind-the-meter gas and nuclear PPAs precisely because the grid is the part they cannot order faster.
Watch next Large-load tariff rulings and 24/7 firm-power deals. If firm power clears at scale, the binding constraint rotates up to compute delivery.
Raw chip supply is largely a procurement problem, not a fab-output problem — Jensen Huang’s own framing. The live constraint is turning a hundred thousand accelerators into one computer: ring-collective topologies stall near that scale, and the next networking jump (co-packaged optics) is ceilinged by indium phosphide laser capacity that sits with a handful of vendors.
Watch next Indium phosphide / EML laser capacity adds, and whether optical-circuit-switched fabrics spread beyond Google. Either eases the interconnect ceiling.
Frontier gains increasingly come from post-training and test-time compute rather than raw parameter count, which changes what is scarce: high-quality RL environments and verifiable rewards, not just FLOPs. The economic constraint is shifting toward cost per useful task — dollars per solved problem, not dollars per token.
Watch next Whether the next generation’s edge comes from a bigger base model or from RL / agentic post-training. The answer reprices the whole model layer.
For most software products the model is now good enough; the binding constraint on who captures the value is distribution, proprietary workflow data, and switching costs. A thin wrapper dies and an embedded agent inside an existing system of record sticks. The edge accrues to whoever owns the last mile to the user, not whoever has the best base model.
Watch next Whether incumbents’ distribution advantage outruns startups’ velocity advantage as agents get more autonomous.
Embodied AI is gated by the cost of collecting real-world interaction data and by the reliability bar of acting in an environment that pushes back — not by raw model capability. Until data collection and sim-to-real transfer get cheap, physical AI scales through narrow, high-value niches rather than general-purpose humanoids.
Watch next Breakthroughs in world models and low-cost teleoperation data pipelines. Either moves physical AI from latent to rotating-in.
Non-model launches that change a lab's capability surface: Claude Code, Computer Use, MCP, Operator, Atlas, Agentspace, Project Mariner. Tracking these separately from model versions because the surface change often precedes the next model bump.
First-class capability packs that the model can invoke autonomously, structured to compose into longer agentic workflows.
Meta acquired a 49% non-voting stake in Scale AI for $14.8B to secure a 5-year data deal (at least $450M/year spend), with Scale founder Alexandr Wang joining Meta to lead Meta Superintelligence Labs (MSL). Rivals Google and OpenAI began winding down Scale AI usage over security concerns.
OpenAI re-anchored the Codex brand on an autonomous coding agent (cloud-and-CLI), competing with Claude Code and Cursor.
Anthropic shipped a first-party agentic coding CLI alongside Claude 3.7 Sonnet, putting Anthropic into direct competition with Cursor and the OpenAI Codex line at the developer-tool layer.
OpenAI's first general-purpose computer-use agent product. Browser-based, available on the Pro tier.
OpenAI's browser product, positioning the chat surface as the navigation layer of the web.
xAI shipped a standalone consumer app and deepened Grok's integration with the X platform.
Google's agent platform layer for enterprises, integrating Gemini with Google Workspace and third-party data.
Google's research preview of a Chrome-integrated browser agent built on Gemini 2.0.
Anthropic published MCP as an open specification for connecting models to tools, data sources, and external systems. Adopted broadly across the field within months.
Apple shipped on-device AI features and Siri's first material capability upgrade in years, with a planned ChatGPT bridge for harder tasks.
Claude 3.5 Sonnet's computer-use API gave the model a screen-and-keyboard surface, opening the path to general-purpose desktop agents.
Microsoft restructured Copilot around a more conversational consumer surface, with voice and persistent memory.
Enterprise tier with expanded context window and admin controls; the formal commercial product layer above Claude.ai Pro.
Persistent context buckets in the Claude.ai consumer surface, the consumer-facing analog of the team-knowledge pattern.
Google's enterprise developer assistant, integrated with the GCP and Workspace surfaces.
Google rebranded Bard to Gemini and launched the Advanced paid tier on top of the rebrand.
User-built configurations of ChatGPT, the first attempt at an app-store layer above a foundation model.
Meta launched its consumer AI assistant across Instagram, WhatsApp, Messenger, and Facebook with Llama-family backing.
The launch that started the modern AI consumer category. Free, conversational, and on the web.
A dated read, not advice. Every claim links to its source on the site, and each card names the single signal that would change it. Updated by hand as the stack moves; useful for orientation, dangerous as false precision.