Anthropic IPO, lessons from Wall Street Technology

The Anthropic IPO, what it means, and whether it will ever happen continues to grab a lot of mindshare. A case of all publicity is good publicity, or a train wreck unfolding in plain sight?

I’ve no idea, only time will tell.

My take on Anthropic, the IPO, and the probability that they survive and take over the world is probably a bit more nuanced than most of the commentary.

I tuned into the B2G podcast for the 18 months up to the famed Halloween 2025 special episode with Sam Altman and Satya Nadella. The petulant novice CEO facing off against the old master, one of the best tech CEOs the world has ever seen.

The B2G thesis on both OpenAI and Anthropic was a simple extrapolation of their observations of the rise of Google et al back in the day. The core assumption, the firms that get road position at the start of the AI era would go on to hold that road position in the long term. Road position being daily active users, eyeballs on the chatbot.

That assumption on holding road position leading to long term compound growth provides the basis for valuation and equally important the long term committed spend, all $gazillion of it.

There was also a recognition that there would be something of a survivor series with big casualties. In the simplest case, at the time OpenAI, Google, X and Meta were all in the advertising business. The pool of B2C digital advertising $dollars is somewhat finite, so not all of them could dominate that business and survive. Meta and Google had road position with formidable franchises. The implicit assumption was that OpenAI, the new era disrupter could, or would, leapfrog them and take the market.

Anthropic was already focussed on enterprise B2B, but the same principle applies. Overall enterprise tech spend is bounded, if not as tightly as the new technology displaces workers. I say "not as tightly" because there is a constraint on the system, ultimately workers need jobs to pay for the products that the Fortune 500 and all businesses produce. Also, replacing all the workers in the world takes time. In any case, Anthropic are not just competing with every worker in the developed world, they’re also competing with every company in the Fortune 500, not to mention the established B2B tech providers, Microsoft, Amazon Web Services (AWS), etc. A lot of very well run tech firms have their eye on the same pie. Jevon’s paradox is a thing, but how quickly will it manifest in practice?

There is nothing surprising in this analysis. I’m not the first to lay it out.

Wall Street Technology Refresh Cycle

Drilling into the next level of detail. I’ve spent my career in Wall Street front office technology.

The dominant feature of that sector before the Credit Crunch was rapid innovation and the three year technology refresh cycle. In that era the global banks were the formidable innovation engines. Running vendors in that environment, back when deals were one-time plus maintenance meant one thing, we had to evolve our products every three years to stay at the edge, to be able to sell new technology into the next refresh cycle. We didn’t run software shops, we ran innovation shops. Loose touch with the edge, stagnation and death. Front office software vendors had a ridiculous fatality rate. To dominate firms had to learn to innovate at the refresh cycle rate. In that light, the current tech leaders, NVIDIA, Tesla SpaceX, Google, Meta, and AWS have been masters of harnessing the tech evolution cycles for decades. They are not static businesses, they are innovation machines.

With AI that refresh cycle time looks to be months not years. It will likely tail off a bit, but the structure or design pattern, using the AI itself to drive innovation, likely means rapid evolution will remain a feature for the foreseeable, the next 5-10-20 years.

What does this mean in practice?

It is always useful to start with a unit example. The last twelve months saw the emerge of a once in a generation breakthrough product, the harness, Claude Code. Agentic harnesses will change the world.

When Claude Code was first released the impact was breathtaking. The problem of course, if AI is great at one thing, it is great at decomposing and replicating software, only better. Within a few weeks a rash of harnesses appeared, immediately leap frogging Claude Code capabilities. At Options Technology the developers used Claude Code immediately in their personal labs and loved it, but at the corporate level we went straight to OpenCode and didn’t miss a beat. In the AI era, there is no software moat, breakthrough innovations will be replicated in a blink. OpenCode are now past 20 million active users, a huge chunk of the AI developer market.

Doesn’t feel like a market with durable road position that lasts decades?

Meta Muse is a great example. Meta leveraged Claude and Claude Code wholesale to build Muse. Once Muse was built they didn’t need it so cancelled the subscription.

Vendor lock in, interchangeability and APIs

Another feature of the Wall Street innovation phase was the battle between vendor lock in and APIs. For enterprise AI, the firms early in the game need to lock in the clients on the one hand, on the other, the savvy clients will build agnostic API layers and work with multiple vendors from the outset. Get stung once and the client is switched on the second time around.

For a unit example, with market data systems built against the original Reuters APIs from the 1990s have proven very sticky. TREP is still a thing, same for symbology, Bloomberg, or RICs (Reuters Identification Codes). In the later phases, starting with TIBCO, then post 2000 the game changed to agnostic APIs that allowed applications to interface with any underlying data source or symbology. This was a big part of the Wombat Financial Software business at the time. Clients are using the AtlasFeed TREP adapter to facilitate vendor switching today. It is worth pointing out that a key difference between the market data business and AI is that with market data the content itself is the key differentiator. Bloomberg, LSEG Refinitv, etc, are content firms. Content is their moat, the distribution layer is secondary.

The API analogy looks horrible when appleid to OpenAI and Anthropic. Both model families are accessed through open APIs, access to the models was fully interchangeable from the outset. This interchangeability even applies at the unit level with Claude Code. If a developer loves Claude Code it’s trivial to point the harness at non Anthropic models, frontier or open weight. Switching is trivial.

Doesn’t bode well for holding road position for a decade.

Don't compete with your customers, or condescend

The next aspect of Wall Street technology back in the day was that the trading firms had a deep seated distrust, or even hatred of two groups of people. The first, anyone who was inclined to compete with them in their core business. The golden rule for vendors, don’t compete with the clients. The second, while they loved working with smart people, talent they could leverage to further their business objectives, they hated being lectured or told what to do by delegations from Silicon Valley. People who thought they were smarter, on a different level to their dumb Wall Street cousins. The Wall Streeters were very tuned in to even a hint of a conflict of interest or irreverence. Rightly so.

Anthropic’s strategy punches both these principles on the nose. They rotated through one sector at a time, finance, legal, security, biotech, gobbling up IP and releasing products designed to eat a huge chunk of business from each of these sectors. Similar vein, both Anthropic and OpenAI launched systems integrators to compete with that ecosystem. Their objective is to compete with all their enterprise clients. Similar for hubris, their entire brand is based on dominating the world.

Apply the lessons from Wall Street, one view is that the business world at large leapt in and leveraged what was breakthrough technology at the time. There could be an equally fast exodus once neutral alternatives appear. The market will gravitate towards providers who are trying to enable their clients, not compete with them.

Once again, doesn’t bode well for durable road position.

Envy, hatred and brutal competition

Aristotle, Adam Smith and Peter Drucker all agreed on one core principle of envy and competition, also core to Wall Street: "potters envy potters."

The bitter rivalry between the young Bill Gates and Larry Ellison in the systems era embodied this principle. Their outright hatred of each other drove competition and rivalry for decades.

Applying today, the inference is that the older generation of tech titans, Huang, Musk, Bezos, Zuckerberg, Karp, Nadella, Brin, entrepreneurs who grafted for decades, hate Sam and Dario with a passion and will do anything within their power to make them fail. The same applies to the corporate identity. Microsoft likely hates OpenAI at a level they’d never admit. Nets to a large group of talented and motivated people and teams hell bent on assuring their failure. Sergey Brin didn't come back to the office because he needs the money. He's driven by something more fundamental. Likely great for innovation, but once again doesn’t speak to durable competitive advantage.

The run up to the SpaceX IPO might provide a useful insight into this dynamic. One common view is that Musk secured road position and viability, and might even have blocked one or both of Anthropic and OpenAIs IPOs by getting out first. OpenAI was arguably best placed for a quick IPO, but Musk was able to distract and pump some gravel in the gears with the failed legal challenge to the viability of the OpenAI corporate structure. When the final ruling was announced it was clear that Musk’s team would have known they could never win the case. It was already past the statute of limitations.

The question, did Musk know he was blocking the IPO and potentially torpedoing OpenAI below the waterline by taking the challenge?

Warren Buffet encapsulates this principle in the adage, "friends come and go but enemies accumulate." Dario, Sam, OpenAI and Anthropic likely have a lot of powerful enemies.

Spin outs and team hires

A final aspect of the Wall Street ecosystem that likely applies is the “team spin out” and “team hires”. Anthropic formed when Dario et al spun out of OpenAI and started their own firm. OpenAI the corporate effectively spun out of OpenAI the non profit, escaping from Musk in the first place. Both Meta and Google made notable team hires, or acquired companies to get key teams. More Wall Street firms than I care to think about formed in this way. When we joined Wombat in 2003, it was a team hire from First Derivative / KX. We made a team hire from Kainos seeded the original Wombat Belfast development team in 2004. In 2012 we effectively did a team for equity swap to create what is Options Technology today.

What does this mean?

Applying this principle, it is only a matter of time before teams from the model labs spin out of OpenAI and Anthropic and build their own open weight model companies. The models created will leapfrog the capabilities of the models built inside the Frontier labs. Their final gift to the world.

Conclusion

Parsing all of this, the core assumption behind the valuations and forward spend commit is that Anthropic and OpenAI have road position which will prove durable for a decade. This assumption looks shaky at best. Only time will tell.