Where We Go Post-LLM

Hey y’all — where do LLMs go from here?

Way back on February 1st I explained why LLMs were already “over.”

My point was that LLMs were doomed to a race-to-the-bottom in terms of pricing, which would lead value to accrue elsewhere in the stack.

And this is exactly what we’re seeing.

A growing percentage of startups that a16z looks at (16-24% according to partner Martin Casado) use open source Chinese LLMs rather than or in addition to the American brand-name versions.

Presumably any difference in quality is accounted for by robust prompt engineering or, maybe more likely, at the early stages of a new company it (and data security) actually doesn’t matter that much and price is more important.

This week I’m taking a look at some of the logical questions that come up from this trend like:

• Why is this happening?

• What does this mean for OpenAI, Anthropic, etc?

• Where should you be building / where will value accrue?

Where We Go Post-LLM

Why LLMs Are (Still) Over

Let’s briefly play out the relational database example.

In the 90s Oracle, IBM, and Microsoft charged premium prices for relational database management systems.

Eventually, free open-source alternatives like PostgreSQL and MySQL became viable competitors and completely commoditized the market.

As the market evolved, the cost of running software built on top of relational databases went to near-zero and value accrued where differentiation remained — which was in the data itself (the data layer) and how that data was used (the application layer) rather than the database itself (the infra layer).

Today we see the same dynamic playing out:

• GPT-5’s release demonstrated the already-diminishing returns of model quality due to the lack of new data to train on. We’re already at the iPhone 12, not the iPhone 5. Without a true breakthrough, this will continue.

• Open source and Chinese entrants will continue driving pricing down.

• Google is loving this because, due to their cash position, they can outspend the relative upstarts like OpenAI forever — creating challenges for OpenAI to reach their aggressive capital commitments without eventually taking on substantial dilution.

What This Means & What Happens Next

The subtext here is sort of “does this mean OpenAI, et al are doomed?”

The answer is no.

LLMs may not be where the value will accrue over time, but it doesn’t mean building them is pointless or that OpenAI and Anthropic will live and die by the value of LLMs alone.

Anthropic’s strategy appears to be building for the enterprise. They want to deliver the best experience to, and lock in partnerships with, large ACV enterprise players. My guess is they see that as the foundation and anything else as downstream.

They already server over 300,000 enterprise customers and have focused internationally by emphasizing their safety-first approach.

OpenAI’s focus seems to be more about venturing into the application layer, hardware (both in terms of data centers and consumer devices), and other bets themselves.

Their phone “companion” device, designed by Jony Ive, will at least be formally announced next year, their Stargate data center project is reportedly $500B, and their software focus is clearly shifting towards applications and agents around coding and more.

Both strategies are more than just a hedge against the issues with the future of LLMs we’ve noted, but rather the logical next step for companies in their position.

Where You Should Be Building

If LLMs are the new PostgreSQL — ubiquitous, cheap, and increasingly interchangeable, then the real winners for 2026 will be in things like:

• Curating unique data

• Applications that embed AI into sticky, retentive workflows

• Agents that act rather than just chat

Martin Casado from a16z has flagged in recent talks that the AI stack is maturing just like the internet did with commoditized infra at the base and moats higher up at the data and application layers. The difference this time is that there will be a third valuable layer: agents.

Target verticals where there’s either a scarcity of data (and you have a way to acquire it) or the workflow is super painful (and LLMs can make it better). Ideally both.

Most importantly, iterate through your ideas quickly. With decreasing model costs, you can afford to.

How We Can Help

Advertise in my newsletter to get in front of 60,000+ founders.

_{Figure Lending LLC dba Figure 650 S. Tryon Street, 8th Floor, Charlotte, NC 28202. 888) 819-6388. NMLS ID 1717824. For licensing information go to www.nmlsconsumeraccess.org. Equal Opportunity Lender}

_{Digital currency is not legal tender, is not backed by the government, and BIA accounts are not subject to FDIC or SIPC protections.}

_{Crypto loans are offered to U.S. borrowers by Figure Lending LLC. This product is not available to U.S. residents of DC, ID, IL, KY, MD, MS, SD, TX, VT, or VA.}

_{Crypto loans are offered through Figure Markets Credit LLC to residents of the state of New York and to international customers except in the following jurisdictions: Crimea (Ukraine), Donetsk (Ukraine), Luhansk (Ukraine), Afghanistan, Albania, Belarus, Central African Republic ,Congo (the Democratic Republic), Cuba, Ethiopia, Haiti, Iran (Islamic Republic of), Iraq, Lebanon, Libya, Mali, Myanmar (Burma), Nicaragua, Nigeria, North Korea (Democratic People's Republic of), Pakistan, Palestine (State of), Russia, Somalia, South Sudan, Sudan, Syria, Ukraine, Venezuela, Yemen, or Zimbabwe.}

_{The interest rate on Figure's Crypto-Backed Loan is 8.91% (9.999% APR) at 50% LTV or 11.50% (12.62% APR) up to 75%. The advertised Annual Percentage Rate (APR) is based on a $5,000 loan amount and contemplates the payment of 1% origination fee, which would yield a monthly payment of $42.47. Rates will be higher for applications secured by assets with a higher LTV ratio. The Figure Crypto-Backed Loan has a 12 month interest-only repayment term and allows for a maximum initial LTV ratio of 75%. Interest rates change frequently so your exact interest rate will depend on the date you apply and may depend on many factors such as LTV ratio.}