Chamath Palihapitiya's AI Investing Framework

LPS stands for Land, Power and Shell — the physical footprint of a data center before any chips go in. The site, its connection to the grid, and the building. Nothing computational about it.

This is the layer Chamath Palihapitiya is most emphatic about. In his August 2026 post he described it as still the most obvious and fastest path to cash on cash returns, adding that lots of value can be assembled and traded quickly at this layer. He has said he holds close to 6GW of power running through 2029.

The argument is about scarcity rather than technology. A GPU can be manufactured. A site that is already energised and permitted cannot be conjured on demand — it is constrained by grid interconnection queues and local politics, both of which move on timescales measured in years.

The corollary he draws is that the bottleneck has moved. When compute is sold out and models are commoditising, the binding constraint is not silicon. It is electricity that exists today.

He has said he will not invest in or incubate anything in the silicon layer, and expects a lot of capital to be wasted there. His stated reasons are that chip performance demands are too high, manufacturing precision is too complex, and a startup can no longer exert the supply-chain influence needed to secure adjacent components such as memory.

This carries weight because of where he is arguing from rather than against. Social Capital was one of Groq's earliest backers, investing $10 million in 2017 and a further $52.3 million the following year, which gave the firm close to a third of the company. In December 2025 Nvidia agreed to license Groq's inference technology in a deal reported at approximately $20.6 billion.

Chamath Palihapitiya has argued that assigning large terminal value to the model layer is “a mathematical mistake”. His reasoning is that frontier capability is being replicated far faster than it used to be: commoditisation that once took five to ten years now takes months, with open-weight releases such as Moonshot AI's Kimi K3 cited as evidence.

In the public portion of his May 2026 essay on the AI stack he put the cost trend plainly: The price of running a model has dropped 1,500x in six years, and intelligence is becoming free.

That is not a claim that model companies are worthless. It is a claim about durability— that a lead measured in months does not support a valuation premised on a lead measured in decades. Which sets up the more uncomfortable question about their revenue.

Tokenmaxxing is Chamath Palihapitiya's term for wasteful internal AI usage — employees consuming large volumes of tokens without a matching return. He raised it on CNBC on July 14, 2026.

The mechanism he describes is an accounting one. The spending is real but diffuse, spread across teams and tools rather than sitting in one line item. His prediction is that it eventually surfaces the hard way — as an earnings miss driven by costs executives didn't know existed inside of their organization.

The investment implication runs in both directions. For companies buying AI, it is an unbudgeted expense. For the labs selling it, some portion of reported revenue may be waste rather than durable demand — which means it can compress even while token volumes rise.

He is not exempting himself. His own company, 8090, was reported in March 2026 to be on track to spend more than $10 million a year on AI usage credits.

An AI harness is the software layer wrapped around a model that decides what the model sees, which tools it can use, and when it stops. The model supplies raw capability; the harness supplies the context, the guardrails and the evaluation.

This is the layer Chamath Palihapitiya says he is building in, and the reason he founded 8090. His argument is that the harness is where a company's own advantage lives — its data, workflows, evaluations and business rules — and that keeping it separate from the model makes the model interchangeable.

Those two sentences carry the whole thesis. The first says the harness is the answer to tokenmaxxing: better orchestration cuts consumption without cutting quality, which is precisely the dynamic that would compress model-layer revenue. The second says the durable value ends up with whoever owns the context, not whoever owns the weights.

8090 raised a $135 million Series A led by Salesforce Ventures in June 2026.

The framework is coherent and specific, which is more than most public commentary offers. It is also, at every layer, aligned with positions he already holds.

None of that makes him wrong. Investors with conviction generally do own what they recommend, and the alternative — commentary from someone with nothing at stake — is not obviously more reliable. But it does mean the framework should be read as a position, not a neutral survey, and the parts that are hardest to verify are the parts where he has the most at stake.

The testable claims are the ones worth tracking. Whether energised power stays scarce, whether open models plus harnesses actually cut token spend, and whether model-layer revenue holds up as that efficiency arrives — those resolve on their own schedule, regardless of who is talking.

The same discipline applies to reading any investor's public position. Disclosure shows what someone holds, never what it cost them or what sits behind it.