IGNITION: Off With Software’s Head!

Welcome to Ignition, Catalyst Investors’ briefing on what we’re seeing at the intersection of AI, robotics, software, and growth equity, combining the best of human and AI expertise.

In this issue:

Google advances methodically in the AI race
Software goes headless, led by Salesforce.com
The sales and revenue AI market map

01 — THE SIGNAL

The Google Tortoise Keeps Advancing

For two years, the dominant narrative on Google and AI was some version of the same anxiety. OpenAI had consumer momentum and Anthropic had business users. Amazon and Microsoft had the compute. Google had everything and somehow seemed to be losing anyway. The company that invented the Tensor Processing Unit, that owns DeepMind, that has more AI infrastructure than anyone on earth, plodding along while the others are off to the races.

The Google I/O event reframed that story with a picture of what it means to own the full stack.

The Model is Not the Product

Every other company in this race is selling intelligence. Google is selling runtime.

The distinction matters. Gemini 3.5 Flash, the centerpiece of I/O, is not Google’s answer to Claude Opus or GPT-5. It is deliberately positioned below the frontier on raw model quality and priced accordingly, at $1.50 per million input tokens, roughly half of Claude Sonnet. Google is betting that frontier-quality reasoning is fast becoming a commodity, and that the race that matters is to control the tracks where AI takes action.

Sundar Pichai made the cost argument explicit on stage…enterprises running heavy workloads on frontier models from other labs could save over a billion dollars annually by shifting 80% of those workloads to 3.5 Flash. Independent testing by Artificial Analysis suggests quality is competitive on speed-sensitive tasks while trailing Claude on complex reasoning and coding. The user reviews are mixed. One Uber ML engineer on Google’s developer advisory board said that the cost savings are not worth the quality gap. On the other hand, a non-technical startup founder who got early access described Antigravity (Google’s coding agent) as now “close to parity” with Claude Code for her use case. Both assessments are probably true for different customers, which is exactly Google’s market segmentation strategy.

The full Pro model, Gemini 3.5 Pro, is in internal testing and ships next month. That is when the quality argument will get properly stress tested.

What Google Has That Nobody Else Does

The more important story at I/O was the agent infrastructure built on top of the model, and the ecosystem that infrastructure runs on.

Gemini Spark is a 24/7 personal AI agent running on its own Google Cloud VM, with native access to Gmail, Calendar, Drive, Docs, Sheets, Maps, and YouTube. Antigravity 2.0 is the enterprise and developer version of the same idea: parallel agents, background tasks, full integration with Google Cloud’s data residency controls. Search was rebuilt for the first time in 25 years to accept text, images, files, video, and Chrome tabs as input, and now supports agentic booking for local services. Smart glasses ship this fall.

The through-line across all of it is the single architectural insight that agents don’t need better models, they need more surfaces to act on. An agent that can read your inbox, check your calendar, search the web, process a payment, and route a result to your phone is structurally more valuable than ten specialized agents that each do one of those things in isolation. Google can give its agents that full surface area under a single identity layer. No one else can.

Consider the competitive landscape. Microsoft has strong cloud infrastructure at Azure and model capability through its OpenAI partnership, but its consumer AI footprint is thin, and its advertising business is a fraction of Google’s scale. Amazon owns cloud infrastructure at AWS but has no meaningful AI interface beyond Alexa. Meta commands advertising reach that rivals Google’s but has no cloud business, no enterprise software, and its best model play remains Llama, an open-source offering that trails the frontier labs. OpenAI has the best consumer brand and a credible argument for leading model quality, but no built-in distribution infrastructure and expensive compute commitments. Anthropic has the strongest position in enterprise coding and reasoning, but is, as we noted in Issue #1, renting its compute from a direct competitor.

Google’s Q1 numbers tell the underlying story. Search revenue grew 19% year over year to $60.4 billion, directly contradicting the bear thesis that AI chatbots would cannibalize search advertising. Google Cloud grew 63% to $20 billion in the quarter, now generating $80 billion in annualized revenue and growing faster on a percentage basis than AWS. AI Overviews, the generative answer panels embedded in search results, are monetizing at rates comparable to traditional search. The cannibalization scenario that defined the Google bear case for three years is not happening.

The Honest Counterargument

Google’s quality gap in coding is real. Developers choose tools on performance, not price, and Claude and Codex currently have the better performance for serious engineering work. Google’s strategy of winning on cost and surface area only works if 3.5 Pro, when it ships, closes enough of that gap to make the ecosystem advantages decisive. And lock-in cuts both ways: the same integration depth that makes Spark compelling for existing Google Workspace customers makes it irrelevant for organizations committed to Microsoft 365 or deep in Anthropic’s API. Google is not winning those customers back with a better Flash model.

The Frame That Matters for Investors

While the hares have been running faster, the tortoise has been building roads.

What I/O made clear is that the roads are now paved enough to matter. Google did not win the model race in the last two years, but it did build the AI infrastructure layer. That infrastructure is now generating 63% cloud growth, monetizing at search scale, and expanding into agentic AI with distribution that no competitor can match.

The investment question is not whether Google has caught up on models. It is whether ecosystem and distribution depth compound faster than model quality in the agentic era. Google is making a large, well-financed bet that they do.

Sources: The Pareto Investor (May 28, 2026), The Information / Erin Woo (May 20, 2026), Stephen Smith / smithstephen.com (May 19, 2026). Q1 2026 earnings figures from Alphabet investor relations.

02 — STRATEGY CORNER

Off with Software’s Head!

Andreessen Horowitz partner Seema Amble published a sharp read on the state of enterprise software. Read it in full: Is Software Losing Its Head?

The prompt is Salesforce’s announcement that it would open its APIs and reposition as a “headless” product, betting that in an agentic world value lies in the data layer and not the UI. Amble asks the next question: when agents bypass the interface entirely, what makes a system of record defensible? She reframes how we think about moats in enterprise software, and it has direct implications for investments in the world of B2B AI.

The agentic era changes how software is used. The moats that made SaaS companies durable for two decades were largely based on human path dependence: training, reporting, the friction of asking a team to learn a new interface. Agents don’t have muscle memory, however.

The key takeaways:

The UI moat is fading. Agents read and write directly to underlying data, bypassing the interface, and erode the stickiness that comes from human workflow habits. Sales reps brought Salesforce experience with them to new jobs. Agents do not have that loyalty.
Operational logic and compliance remain durable. What agents cannot easily replicate is the institutional knowledge baked into workflow rules, approval hierarchies, permissioning structures, and compliance requirements built up over years. That context is exactly what agents need to operate safely, and it does not export cleanly.
Connectivity is harder to unwind than ever. The more a system mediates interactions across functions or external parties like buyers, auditors, or regulators, the harder it is to displace. In an agentic world, a CRM agent can stitch together data across sales, billing, and customer success in ways a human user never could.
Proprietary data exhaust is the new moat. Defensible data is not what you import. It is what your product uniquely creates, like observed behavior, response rates, process outcomes, and exception patterns. Systems that generate new data by being inside the action loop are structurally harder to replace than those that warehouse outside data.
Real-world execution creates a different class of defensibility. Software that closes the loop into services, logistics, field operations, or payments is not just storing a record. It is dispatching people and moving goods. That is a moat no API can easily replicate.

The Catalyst take: we are looking for B2B AI companies that combine at least two of these factors. Compliance-critical data plus real-world execution. Proprietary data exhaust plus multi-party network effects. Powering vertically integrated services and/or physical AI. Single-factor moats that relied on UI stickiness alone are the most exposed to agentic disruption, and the most interesting investment opportunities right now are companies building in categories where the incumbent ends up headless, revealing how thin its moat has suddenly become.

What does this mean for Salesforce.com? Hard to know yet, other than they are on defense. They are the system of record, have years of data, own daily employee communications with Slack, but are vulnerable to an AI native CRM or to vertically native CRMs. When software coding becomes abundant, everything is once again up for grabs. Salesforce starts with an advantage, but it will have to play the right cards to win.