אִם יִרְצֶה הַשֵּׁם
Follow-up to The Collapse of the Abstraction Layer, February 9, 2026
The companies that adapt will not do so by adding AI features. That is the wrong frame. They will do so by asking what they have that an agent cannot assemble from Stripe, Clerk, Turbopuffer, and a few API calls. If the answer is "not much," the strategic conversation needs to start from there, not from the product roadmap.
The new stack is compute as sovereign territory with known power sources, software as permissioned agent protocols rather than interfaces, security as behavioral attestation rather than perimeter defense, and memory allocation as the primary scarce resource. That is a description of what is already being built by the organizations with the runway and the clarity to see it.
>The $300 billion dislocation is the market writing off the old abstractions. The question is what gets built in the space they leave behind.
In February it was argued that the abstraction layer holding together the software economy was under structural stress. Not hype stress. Not "AI is coming" stress. The kind that shows up in valuations, in hiring freezes, in the quiet cancellation of contracts that used to auto-renew.
Most underestimated the speed.
The core claim was simple: AI was quietly dismantling the abstraction layer between "software" and "outcome". SaaS vendors sold workflow wrapped in GUI. The GUI was the product. Once an agent can navigate the GUI, manipulate files, and chain steps without supervision, the GUI stops being the product. It becomes a target.
What few anticipated was how fast the market would price this in, or how many independent observers would arrive at the same structure from different entry points. Neevash Ramdial wrote a sharp piece in May naming four simultaneous collapses: the software stack, the global supply chain, the enterprise/consumer security boundary, and the cloud/physical compute distinction. Different angle, same underlying geometry.
When multiple analysts independently converge on the same failure mode (look at Whewell's Consilience of Induction), that is worth paying attention to.
The 0.5 Layer Is Now Visible to Everyone
In February I framed this as a coming threat. By May it reads as an operating reality. The companies pulling away right now share one property: their products are consumed by software, not by people. Turbopuffer at $100M ARR, profitable, on less than $1M raised total. Modal closing a $355M Series C. Mintlify reporting that AI agents now account for nearly half of all traffic across their documentation sites.
None of them pivoted. None of them rebranded as AI-native. They were already clean enough, fast enough, and API-first enough that a non-human caller could pick them up without help. The agent wave did not create their product. It added a second audience to it.
I called this the primitives layer in February. The sharper name is the 0.5 layer: the auth, vector search, inference, and payment infrastructure that an agent reaches for when it assembles software. The lego pieces, if you will.
The application layer built on top is now the thing an agent can reconstruct in hours from those pieces. That is the structural problem for most of SaaS.
Four Collapses I Missed or Underweighted
Supply chain sovereignty. I focused on the software stack in February. What I did not foreground was that the cloud abstraction itself is dissolving into physical reality. New York moratoriums on datacenter permits. Virginia and Vermont considering similar moves. The Trump administration clearing Nvidia H200 exports to China under a 50% volume cap and a 25% revenue tithe to the US government. Chips are now geopolitical leverage. "The cloud" as a placeless compute substrate is over as a concept.
HBM as the binding constraint. SK Hynix and Micron are sold out of High Bandwidth Memory through 2026. Prices have doubled. Unlike GPU capacity, you cannot rent HBM elastically; you sign purchase agreements 18 months in advance. The hyperscalers saw this coming and locked in allocations. Startups that did not are now structurally disadvantaged in ways that no amount of clever engineering fixes. The scarce resource turned out to be physical memory, not model quality.
Shadow infrastructure, not shadow IT. The old threat model was shadow IT: employees installing unapproved apps. The new one is worse. OpenClaw, the open-source personal agent, is running on corporate machines at 22% of monitored organizations. Cisco found that 26% of agent skills contain vulnerabilities, with 1,800 exposed instances leaking API keys. When an agent treats the browser and command line as its operating system, the VPN perimeter is theater.
The implementation gap. The technology curve and the adoption curve are not close. They are years apart. Insurers currently refuse to underwrite AI risk, which means legal teams are killing productivity projects that would otherwise ship. CFOs see "hours saved" but cannot measure it against headcount. The organizations that most need to adapt are structurally the least able to do so quickly. An 18-month deployment cycle for a properly audited agentic framework is not slow; it is realistic. The frontier moves every three months.
The Question I Asked in February, Sharpened
In February I asked whether your product could survive being consumed by something that does not read landing pages. The sharper version, four months later, is whether an agent given your product's core task would route around you entirely and assemble the same outcome from 0.5 layer primitives.
Not "would it use AI." Not "would it find value." Would it need you at all.
The honest categories of "yes, still needed" are narrower than most companies want to admit: proprietary data that took years to accumulate, regulatory positioning expensive to replicate, trust relationships spanning decades, or genuinely hard technical problems that cannot be vibe-coded in a weekend. Those moats hold. Everything else is a workflow on top of commodity primitives, and that category is the largest single bucket in enterprise SaaS.
Swan Song
The companies that adapt will not do so by adding AI features. That is the wrong frame. They will do so by asking what they have that an agent cannot assemble from Stripe, Clerk, Turbopuffer, and a few API calls. If the answer is "not much," the strategic conversation needs to start from there, not from the product roadmap.
The new stack is compute as sovereign territory with known power sources, software as permissioned agent protocols rather than interfaces, security as behavioral attestation rather than perimeter defense, and memory allocation as the primary scarce resource. That is a description of what is already being built by the organizations with the runway and the clarity to see it.
The $300 billion dislocation is the market writing off the old abstractions. The question is what gets built in the space they leave behind.
Further reading: - Neevash Ramdial, "The AI Bifurcation of Tech: Why the fundamentals matter more than ever", May 2026 - Dave Shapiro, "The Depressing Reality of AI Adoption" - Nathan Benaich, State of AI: February 2026 newsletter