Founder of Social Capital: When Labs Can Produce Similar AI Models, Real Monetization Advantage Comes from Unique Ingredients
Chamas Palihapitiya stated that when labs can produce similar models, the real monetization advantage comes from that unique ingredient that others do not have. If the same thousand inputs are given to Facebook, Microsoft, Google, and Amazon, they will produce similar types of machine learning models; adding one unique ingredient will lead to significantly different outputs, like three chefs using three ingredients, with only one having access to a fourth.
He believes that currently everyone is still crawling public web pages. The next phase will see someone lock down a certain site or type of corpus, only feeding it to their own models, which will perform better. This will turn into an arms race. The next wave of mergers and acquisitions may involve Google, Microsoft, and Facebook examining a batch of companies and only asking one question: Can they provide effective input for my large model or other machine learning systems?
In the same segment, he compared large models to refrigeration technology: the inventor of refrigeration made some money, but it was Coca-Cola that built an empire using refrigeration. Models will make money, but Coca-Cola has not yet been built. Therefore, early-stage funds can invest in companies that have almost no prospects for independent listing but have built a unique data warehouse, as that warehouse will be bought by larger players to feed models.
This idea first appeared in his interview about asset classes where only about 10% of institutions can truly make money, and it has been circulated again in recent years. He later referred to the terminal value of models as a mathematical error, stating that intelligence is becoming cheaper, and he has positioned himself in land, power plants, harnesses, and application layers, founding 8090 to create a layer that retains proprietary context. Nadella has also publicly written: You pay for intelligence twice, once in cash and once with proprietary knowledge that must be fed in to make the model useful.
In market mechanisms, buyers are large-scale platforms seeking differentiated outputs, while sellers hold workflows, transactions, medical records, industrial logs, legal documents, design files, or user behaviors, but may not be able to grow into independent public companies. The driving force is the homogenization that occurs after public web pages have been crawled. Funds are shifting from "retraining a similar base" to "buying a unique ingredient." Those who can turn boring private data into compound inputs benefit, as do platforms that lock ingredients into their own models through mergers and acquisitions; those relying solely on public corpora for fine-tuning, without a fourth ingredient, are under pressure.
Source: Public Information
ABAB AI Insight
Chamas, at Facebook, turned news sources into large-scale online machine learning, knowing that "the same set of features fed into four companies will converge to the same type of sorter." He now elevates this lab knowledge to the industry level: weights can be replicated, but corpus contracts cannot. The refrigeration/Coca-Cola analogy is his asset allocation command to LPs—do not compress the terminal value into the compressor, but allocate it to those with formulas and channels who can turn refrigeration into shelf monopolies. The investments in 8090 and harness are the practical application of the same judgment: separating the company's own context, tools, and downtime conditions from the model, as models are replaceable, but ingredients are not.
The capital path is redefining startups from "future public companies" to "future training inputs." Checks no longer require independent revenue curves to complete an IPO, but only require data flows in a certain vertical that cannot be replaced by Common Crawl. The logic of mergers and acquisitions on the platform side has shifted from buying revenue and users to buying closed corpora and closed workflows, as the marginal information of public web pages is decreasing, and export controls and site anti-crawling further fragment the remaining public sources. Salesforce's acquisition of Informatica, IBM's acquisition of Confluent, and Meta's investment in Scale are all early moves in the same direction.
The analogy is that Google's acquisition of YouTube was not about buying a video site but acquiring a segment of user-generated corpus that cannot be replicated; Bloomberg terminals do not sell news but sell keys and behaviors on those keys; Palantir locks customer workflows into forward-deployed engineers. The industry position has shifted from "who can train a larger base first" to "who can lock the fourth ingredient into the warehouse first." Open-source weights have accelerated this transition: the more models resemble infrastructure, the more private inputs resemble oil fields.
Structural judgments belong to the reconstruction of the industry chain combined with the transfer of pricing power. The mechanism is: when the fitting function is open to everyone, the only difference left is the training distribution. Whoever can shift the distribution from the public domain to the private domain regains pricing power over outputs. Mergers and acquisitions are not just embellishments of exit channels; they are the conveyor belt that moves ingredients from startup balance sheets into platform model cards.
ABAB News · Cognitive Laws
- Models can be the same, but the additional ingredient cannot be the same.
- After public web pages have been crawled, advantages are generated by locking doors.
- It is not necessary to become a great public company; being the missing ingredient in someone else's model is sufficient.