Bitter Lessons From HALO'd Companies
At moments of early fervor, startups emerge who feel they must scale aggressively in order to catch up to early movers. HALO'd companies are prime examples of the downside of this in AI.
AI acqui-hires/Hire-And-License-Out (HALOs) progressed from mechanisms to acquire talent due to talent dispersion (2023) to talent specialty (2024/2025) to talent competition (2025) as Zuck woke up one morning and decided to go scorched earth.1
There has been lot of ink spilled around these dynamics, perhaps best outlined by Kwok and his take that HALOs are in some ways an explicit recognition surrounding talent being worth more than products. This take, paired with the increasing thesis that software itself will have minimal moats in a “prompt to app” post-AI world, actually point to the value of brand moats in AI.
In one case the brand being tied to talent and in the other case the brand being tied to a given product.2
Beneath the meta-games of HALOs there is a less talked about dynamic of the difficulties in building in AI over the past few years that the HALO’d companies (Inflection, Adept, Character, Windsurf)3 are good examples of.
HALOs and Strategic Positioning
If one were to analyze most of these companies,4 the core insight might be that they found themselves strategically mispositioned, in quite short order, relative to founding/funding ideals.
Many HALO’d companies were predicated on a world view of where AI was in the moment without the prescience of properly modeling where the world was going as the AI industry progressed. I outlined this difficulty in my 2023 essay The Dark Forest of R&D and Capital Deployment in AI.
Specifically, I said:
AI companies today find themselves in multiple cycles of R&D and commercialization, making an implicit bet that the millions or billions of dollars spent on R&D will eventually lead to market domination and massive scale years from now. This pushes these companies into a dark forest they must navigate, while stacking up R&D costs ahead of clear staying power for far longer than the vast majority of software businesses.
Most of the HALO’d companies were a result of the structure of capital markets, with founders using pedigree and investor interest/consensus to their advantage at massive (perhaps unprecedented) capital scale only to realize quite quickly the lack of optionality they created at the company level by piling hundreds of millions of dollars at high valuations ahead of requisite traction/product.5
Many purely underestimated the amount of long-term capital that would be required to compete in frontier LLMs and create any sort of compounding moats.
Inflection
Inflection (the one to kick off the HALO structure) is an illustrative example from the outside looking in.
Inflection was predicated on the idea that there was still room for a generalist frontier AI lab, but with a more consumer-oriented approach focusing on personality and memory.
This was not an absurd thesis in the moments before ChatGPT brand dominance (there were still open questions whether the frontier labs would own more of the stack or just be APIs). In addition, talent and capital accumulation by OpenAI and Anthropic hadn’t fully settled yet, and the industry was still before the release of LLaMA-3 which steamrolled a bunch of other under-resourced but maybe-interesting LLM development companies.
Inflection was led by a well-known founder but one who wasn’t able to gather talent at scale relative to their competitors.6 They however, were able to accumulate a large amount of compute after raising $1.3B (isn’t it cute how this was viewed as a lot of compute back in 2023).
What quickly happened was that even the 22,000 H100s were clearly not going to be enough, and the consumer landscape was ceded at the lab level to ChatGPT, Claude, and to a lesser extent, Character.ai + LLaMA forks on the role-playing/personality side.
Inflection quickly became a company that was unable to play the scale games of their competitors across capital, compute, distribution, and talent, and was candidly knocked a tier down in prestige as an AI lab focusing on frontier AI research.
As Jensen began to become a cult figure and a singular bet on scale became evident, The Bitter Lesson reared its head for the upstart labs like Inflection (and adjacently Adept) and almost all companies were viewed as dead on arrival post-2023 if predicated on transformers and scaling as their path to AGI.
Inflection was higher profile than many others which made it difficult to just bury quietly and give back the money78. Leadership recognized this and engineered a landing that did well for investors, helped Microsoft (their largest funder and beneficiary of inference costs) who at the time was struggling on AI leadership, and created the novel structure we’ve come to know as HALOs.
The window closed before Inflection could even allocate their GPUs properly.
Character.ai
This “not enough scale” path also rings true to Character.ai’s pre-HALO journey.
Seemingly, Noam believed that Character would be a nice mechanism to build a large dataset (and perhaps business to pay for compute) via the chat interface on the role-playing side. Interestingly enough, this is also the same premise that Hugging Face was built on back in the day.
What Character learned over time is that in-context learning allows for some of this behavior to create decent role-playing dynamics at the model level, and the open-source of LLaMA meant that they were once again under-gunned on scale. To actually build true AGI (instead of horny chatbots), the compute and capital would likely accumulate to larger companies instead of Character.
Having a consumer product would not be the best path unless you could monetize it at large scale, and Meta’s push likely meant Character was fighting an uphill battle. With this in mind, if the thing you cared about was AGI, you might as well go back to Google, as Noam did.9
Codeium/Windsurf
After another year or so and many billions of dollars, we then entered into v2 of the HALO paradigm with Windsurf.
Windsurf is/was illustrative of the progression of the space as more narrow companies were deemed valuable-ish once the frontier lab meta had shifted from “API for everything” to “AI for everything”.
Early in this transition/exploration the frontier labs recognized their early verticalization attempts should be towards coding, a problem set they natively understand and which has played well into test-time compute scaling paradigms due to its verifiability.
This direction of capital + compute scale towards coding meant Codeium models and more importantly distribution, were not mid-term going to be able to outpace frontier labs as they shipped Codex/Claude Code and potentially overfit their models towards coding tasks to generate ARR ramp in the short-term to fund AGI ramp in the long-term.
Windsurf/Codeium was prescient early on, being founded somewhat ahead of the AI hype in ~2021. Their HALO was as much about defeat as strategic misposition, as Windsurf had not been able to build the brand moats or distribution advantages versus Cursor and the company was staring at margin compression and second place (at best, likely 4th/5th place if you include other labs) in a power-law and oligopolistic market.
This pressure felt evident as despite the business doing well, the capital needed to compete for the next five years likely would prove untenable, spelling clear difficulty ahead. Thus, the company split, controversially going to Google, with the remaining employees merging with Cognition in an attempt to build a fuller-stack company with a slightly different market positioning than competitors today (a worthy bet to make).10
Kwok’s points are best on this so I’ll just quote him:
Windsurf’s recent acquisition is an interesting example. It had a real product and business with $80M+ ARR. And OpenAI’s original acquisition would have been for the entire business. So what exactly is Google buying by doing a HALO? The thing they care most about capturing is not the current customers or revenue. It’s a team of people who have proven they can identify and organize their efforts in a way that is able to repeatedly translate these models into products customers want. Google, despite its superior models, has not yet shown they can do that–whatever one thinks is the thing getting in the way. The belief that the Windsurf team can do this is the core thing Google is trying to ingest. The details of rebuilding (or building anew) the product or customer base it believes are simpler to be redone.
Moving Forward
As the AI industry meta continues to shift, I imagine there will be newer lessons to learn and perhaps even a new wave of M&A and HALOs that sit outside the frontier labs.
Tech is quite good at knowing or being forced into learning how to bundle and unbundle.
This next wave of consolidation could look like application layer companies merging, infrastructure tooling consolidating into a few players who can actually capture value versus horizontal platforms, and/or specialized hardware plays getting acquired for IP that changes unit economics.
Lessons from Moments of Fervor
While I’m sure there are details I’m missing in my analysis here, the past few years of HALOs are illustrative of a recurring lesson in our increasingly momentum-focused startup ecosystem:
At moments of early fervor, fast-followers emerge who feel they must scale aggressively in order to catch up, often via large capital raises ahead of PMF.
This approach forces companies into a single attempt to unseat incumbents and launch themselves into the echelon of their competitors before investor and talent sentiment wanes. Startups are rarely successful on this type of path.
While all companies have a singular bet the company moment in their journey, most of the time this happens once the company has been able to navigate a variety of idea mazes and mature as an organization. Founders in AI taking this approach are doing the equivalent of a Star Wars Trench Run11: one shot at glory in the infancy of their company, with the other option being death/exile into a Mag7 lab.
This dynamic has incinerated a very large amount of capital and talent hours among the first wave of fast-followers. The lesson in early HALOs is not that their ambition was misplaced, but that their timing and strategic positioning was fatal. In moments of platform shifts, the companies best positioned to win are either the early movers who define the space, or the final movers who arrive after the brutal physics of the market have been established for a final turn of disruption.
For those building in this interim period, it is existential to recognize which games are already over in the in-between periods of these two points.
There’s some comparison to the idea that companies pay people to just sit and not work at competitors.
As we now see with Cursor becoming the latest tech company to be made into a pronoun and a slew of Cursor for X startups coming.
I’m not discussing Scale here as it had actually built a valuable revenue generating business and seemingly was taken out at a price above what many investors/acquirers would have paid for it. Covariant also is not as relevant despite having similar structure.
I'll preface this by saying that in general, I hate when investors speak negatively at earnestly trying startups, especially when they are still private companies and still trying to figure themselves out.
But I guess good optionality at the founder level
This points perhaps to the imbalance of perception between investors and operators/practitioners
Also the founder literally wrote a book about AI, which I said in October 2023 was far too late.
and it was based in the US so we didn’t have the government propping it up like other labs…i know this is ironic as we nationalize Intel…a topic for another day
Character is still one of the more interesting ones to watch as there is real usage on that site, despite the model development no longer being on the frontier
The entire Windsurf story is not too dissimilar to what we saw happen in ride-hailing, both with Lyft (started before Uber) as well as with regional players realizing the costs to compete with at scale players who had large capital cannons were untenable.
Deep cut for those of us on AI twitter