אם ירצה ה׳
TL;DR
We are living through a paradox that would have seemed absurd five years ago: artificial intelligence is advancing faster than at any point in human history, yet deploying it responsibly has never felt more precarious. The frontier models released in January (Claude Opus 4.6 (with its 1M-token context window) and GPT-5.3-Codex dropping within minutes of each other) represent capabilities that dwarf last year’s state-of-the-art. And yet, the $285 billion wiped from software stocks in two weeks tells a different story. Not a story of hype, but of structural collapse.
What connects the market panic, the geopolitical chip wars, the feral agents of Moltbook, and the datacenter moratoriums spreading across US states is singular: the collapse of abstraction layers. The tidy distinctions we built our digital economy upon; between software and infrastructure, between enterprise and consumer, between cloud and physical; are dissolving simultaneously.
We have moved from autocomplete to agents in just a few years, but the organizations tasked with navigating this transition are structurally unprepared for the messiness of implementation. - Dave Shapiro
The Four Collapses
1. Agents Collapse the Software Stack
For two decades, Software-as-a-Service thrived on abstraction. The vendor handled the database, the compute, and the interface; the customer paid per seat for access to a workflow wrapped in a GUI. This abstraction created the predictable recurring revenue that justified SaaS valuations.
The launch of Anthropic’s Claude Cowork in January destroys this layer by inverting the relationship. When an agent can navigate a computer interface, manipulate local files, and execute multi-step workflows autonomously, the "application" becomes merely a target for automation. The value migrates from the interface (the SaaS wrapper) to the orchestration layer (the agent itself).
This is why hedge funds have shorted $24 billion in software names this year alone. Investors are not pricing in a recession; they are pricing in terminal value collapse. If core workflows in legal, finance, and marketing can be rebuilt AI-first at a fraction of the cost, the long-duration revenue streams of traditional SaaS are no longer safe. The abstraction of "software"—the comforting illusion that we buy tools rather than outcomes—has been punctured.
2. Sovereignty Concerns Collapse the Global Supply Chain
The abstraction of "the cloud" as a placeless, frictionless compute substrate is also disintegrating. When New York imposes a three-year moratorium on datacenter permits and states like Virginia and Vermont consider similar legislation due to tripled electricity demand, the physicality of AI infrastructure becomes undeniable.
Simultaneously, the Trump administration's tactical shift to clear Nvidia H200 exports to China—under a 50% volume cap and a 25% revenue tithe to the US government—marks the end of the "global semiconductor market" as a coherent abstraction. Chips are now strategic leverage points. With Chinese customs blocking the first shipments and the Bureau of Industry and Security tightening controls, the abstraction of globalized trade is collapsing into a patchwork of national compute reserves and export-controlled inference capacity.
3. Security Failures Collapse the Boundary Between Enterprise and Consumer AI
Perhaps the most insidious collapse is the dissolution of the perimeter between "enterprise-grade" and "consumer" AI. OpenClaw—the open-source personal agent that became the fastest-growing GitHub repository in history with 157,000 stars—exemplifies this breach. When 22% of employees at monitored organizations run OpenClaw on corporate machines, the abstraction of "shadow IT" mutates into shadow infrastructure.
Cisco's finding that 26% of agent skills contain vulnerabilities, combined with 1,800 exposed instances leaking API keys and credentials, reveals that we have collapsed the boundary between untrusted consumer tools and privileged enterprise environments. The "lethal trifecta"—access to private data, exposure to untrusted content, and external actionability—exists because agents treat the browser and command line as their operating system, rendering traditional VPNs and MDM policies meaningless.
Moltbook's viral moment—where 1.7 million agent accounts supposedly established "Crustafarianism" and debated philosophy—was ultimately debunked as "AI theatre" powered by humans. But the debunking misses the point: we have reached an epistemic threshold where the line between autonomous agent behavior and human performance is genuinely hard to draw. The abstraction of "authenticity" has collapsed.
4. Memory Constraints Collapse the Distinction Between Cloud and Physical Infrastructure
Finally, the abstraction of "cloud compute as infinitely elastic" is running into physical reality. With SK Hynix and Micron fully sold out of HBM through 2026 and HBM prices doubling, we are witnessing a hard constraint that cannot be coded around.
Unlike GPUs, High Bandwidth Memory cannot be rented elastically; it requires long-term purchase agreements signed 18 months in advance. Startups without memory allocation are structurally disadvantaged against hyperscalers like Alphabet and Microsoft, who are doubling their capex to nearly $185B. The cloud abstraction dissolves into a bare-metal scramble for physical memory modules, where the distinction between "cloud native" and "hardware constrained" becomes meaningless.
The Implementation Gap
Here is where Dave Shapiro's analysis becomes essential. Shapiro argues that the "depressing reality" of AI adoption is that the technology curve is years ahead of the adoption curve. While the transition from autocomplete to chatbots took two years, and chatbots to agents took three, the realistic timeline for a Fortune 500 company to deploy a properly audited agentic framework is 18 months.
This gap is filled with friction: - Insurance & Legal: Insurers currently refuse to underwrite AI risk, leading legal teams to kill productivity-boosting projects. - The CFO Hurdle: While technologists see "saved hours," CFOs struggle to measure the ROI of agents that might just lead to "slacking off" once tasks are finished faster. - Labor Displacement: The collapse is already showing in the margins. Beyond the 54,800 AI-attributed layoffs in 2025, there is an "invisible" job gap of 100,000 to 350,000 roles—jobs that were never created or contractors whose terms were simply not renewed.
The New Architecture
What replaces the collapsed layers? Likely a more explicit, less abstracted stack:
- Compute as Sovereign Territory: Not "the cloud" but specific geopolitically-situated clusters with known power sources.
- Software as Agent Protocols: Not applications with interfaces, but permissioned capabilities that agents negotiate access to.
- Security as Behavioral Verification: Not perimeter defense, but continuous attestation of agent behavior (addressing the "nightmare" vulnerabilities found in OpenClaw).
- Memory as the Primary Scarce Resource: Physically reserved HBM allocations that determine who can train and who can merely inference.
The transition will be messy. The $300 billion dislocation in software markets is the recognition that the old abstractions were load-bearing structures that are no longer holding the weight. As Shapiro notes, the companies that survive will not be those that "wait for the hype to die down," but those with the organizational flexibility to operate in a world where the stack has already fallen.
Further reading:
- Dave Shapiro, "The Depressing Reality of AI Adoption"
- Nathan Benaich, State of AI: February 2026 newsletter
- MIT Technology Review, "The Making of Moltbook" , February 2026