Luke Sophinos: The Flip in AI Vertical Applications Has Just Begun
Venture capitalist Luke Sophinos stated that venture capital excels at betting on the frontier, marking the next wave of opportunities. He arranged the history of technology in four phases: first investing in infrastructure, followed by a flip in the application layer. Currently, artificial intelligence is still in the infrastructure stage, and in the next two to three years, many industry vertical application companies will emerge.
His timeline is as follows: during the internet bubble, funds flowed into Cisco's networking equipment and AOL's access, while most people were still using dial-up; then Amazon turned e-commerce into a business, Google made search a money printer, and eBay enabled strangers to trade online, rewriting categories. Before the mobile era, investments were made in broadband and chips prior to the 2007 iPhone; then WhatsApp facilitated cross-border communication, Uber turned smartphones into taxis, and Instagram made photo sharing a global habit, with applications rapidly reaching billion-dollar valuations on smartphones. In the cloud era, investments were first made in developer clouds and databases like DigitalOcean and MongoDB, with AWS leading the infrastructure; then Dropbox enabled file synchronization, Snowflake revamped data warehousing, and Stripe simplified online payments.
In this round of artificial intelligence, he pointed out that funds are flowing into large models and data centers, citing Anthropic, xAI, CoreWeave, and Cerebras as capital-intensive infrastructure. The flip in the application layer "has just begun." In other posts, he completed the logic: if the application layer cannot demonstrate cost reduction, the narrative around infrastructure will become a bubble; value will shift from the interface of seat software to the orchestration layer that executes workflows; vertical industries, due to context, regulation, and service depth, will not be fully consumed by a general operating system.
This narrative aligns with his ongoing promotion of "headless software": old software still exists, but users no longer enter through the original interface; permissions, audits, and interfaces are called by proxies. He advocates for capturing the AI interface in vertical fields, rather than just adding chatbots to old products.
From a market mechanism perspective, this represents a narrative shift in venture capital from computing power bills to industry cash flows. Buyers need to find growth funds that can generate priced revenue after models and cabinets; sellers are still the model labs, GPU clouds, and custom chip manufacturers that are attracting capital. Funds will not immediately leave infrastructure, but new checks will start to demand unit economics in vertical scenarios. Beneficiaries will be teams that can prove real cost savings for hospitals, properties, and law firms; under pressure will be generic shells that only have demos without bills, and old software that sells chat windows by seat. The historical analogy serves as a timeline for LPs: the pipeline has already been laid once, and the next round should buy stores along the pipeline.
Source: Public Information
ABAB AI Insight
Sophinos positions venture capital as "first funding the pipeline, then funding the stores." The four-phase analogy is highly organized: access—e-commerce search, base stations—super applications, cloud—payments and data warehousing, models—vertical agents. This organization itself is a product: it conveniently explains the current positions in fundraising memorandums as "still in the Cisco phase, with Amazon about to emerge." The companies he named, Anthropic, xAI, CoreWeave, and Cerebras, are indeed the most capital-attractive infrastructure symbols for 2025 to 2026; writing them as necessary but not sufficient clears the theme for the next vertical fund.
The capital path shifts from a capital expenditure cycle to a labor substitution cycle. Data centers require electricity, chips, and long-term cloud contracts; vertical applications need industry labor hours and compliance responsibilities. If applications cannot reduce costs for a hospitalization, a work order, or an account reconciliation, the upstream cabinet utilization will lose its demand anchor. He thus writes "measurable cost reduction" as the test for whether infrastructure is a bubble. Money from orders for Nvidia and hyperscale clouds will be partially allocated to teams that can embed into clinics, construction sites, and law firm systems. The interface war will follow: whoever controls the entry point for user intent will call upon whose headless software.
This parallels Marc Andreessen's "software is eating the world" and Benedict Evans' layered diagram: after each layer matures, profits migrate to the next layer's user entry. The difference this time is that the entry point may not be a screen, but an agent. Salesforce interfaces capabilities, and Microsoft bundles agents with workflows, indicating that old system record keepers are also competing for the same layer. The industry stage is a positioning race on the eve of expansion: infrastructure is already expensive, vertical companies are mostly not yet grown, and the narrative first secures the position.
The structural judgment is a reconstruction of the industry chain. The mechanism shifts value from "people clicking in software" to "agents executing between systems." Pricing power is partially ceded from labs with the largest models to applications with industry workflows and regulatory permissions. Pipeline companies become wealthy first, and stores become wealthy later—provided that the stores actually open.
ABAB News · Cognitive Laws
- Venture capital first funds the pipeline, then positions the store as the theme for the next fund.
- If the application layer cannot demonstrate cost reduction, cabinet orders are candidates for bubbles.
- Software can leave the body, but the interface will first lose its head.