Open Models Will Reach Frontier Level in Months | Nick Khami artwork

Open Models Will Reach Frontier Level in Months | Nick Khami

MTS

November 14, 2026

Nick Khami of Morph breaks down the economics of AI compute, debunking the idea that there is infinite demand for raw inference.

0:00

Nick Khami

As the open-source models start to reach, probably, the current frontier capability in the next few months, it gets riskier. Someone says, hey, I'm selling DeepSeek for 40% off of list, and then that sounds like a great deal, you wanna buy it for 40% off list, but how could you possibly know whether it's real DeepSeek or quantized DeepSeek, or maybe the KV cache should just extra compress, so you have some quality discrepancy there, and I guess techniques like this variable inference would let you be confident that you're actually buying what you said, what the seller is saying they're selling.

0:39

SPEAKER_2

We are back, we are live with Nick Khami, who is a member of technical staff at Morph, the inference company for coding agents. Nick, welcome.

0:47

Nick Khami

Yeah, thanks man, happy to be here.

0:49

SPEAKER_2

So first of all, tell us about what Morph is, what you guys do, and what you just announced today.

0:53

Nick Khami

Yeah, so Morph is an AI inference provider. We serve inference on OpenRouter and direct to customers for models like GLM, Kimi, and DeepSeq. Today, we announced Morph Ultra. It's our coding sub, so customers can pay 200 bucks a month to us and get a similar multiple on their usage that they would from Anthropic or OpenAI and it's been cool. I think that we've sold like 80 of the subs since we launched this morning, which is a pretty fast rate.

1:23

SPEAKER_2

Yeah, it's pretty impressive. We were just talking about a few different things before we came on.

I guess we could start with Anthropic last week, just announced progress on formally verifying inference. So they're using some cryptographic technology to allow users or developers to know that they are getting a specific model when they see an output. So you can take a text output and determine provably that it came from a specific model without actually revealing the weights of the model. So do you have any takes on this? How easy do you think it's going to be to put this into production? And what might it be useful for?

2:07

Nick Khami

Yeah, it's like proof of intelligence.

I think a lot of things probably end up getting back by proof of intelligence. So what would be a proof of work, a proof of mistake? Back in the old crypto days, you probably have some task that you want to get done, and you are hoping to delegate it. And today, you're just going on a trust system. So you send it to Anthropic, and you say that you want Opus, and you get back Opus. But as the open-source models start to reach probably the current frontier capability in the next few months, it gets riskier. Someone says, like, hey, I'm selling DeepSeek for 40% off of list. And then that sounds like a great deal. You want to buy it or 40% off list. But how could you possibly know whether it's real DeepSeek or quantized DeepSeek or, you know, maybe the KV cash is just extra compressed. So you have some quality discrepancy there. And I guess like techniques like this verifiable inference would let you be confident that you're actually buying what you said, what the seller is saying they're selling.

Which I think makes it easier to build token marketplaces. It probably makes it makes it more makes inference more fungible as a product. Whereas today, if you're buying from more first, you're buying from like a random company.

Let's say Base 10 versus Morph. Base 10 maybe is a stronger brand than we'd even been around for longer. There's some trust factor associated with that brand that Base 10 has built. And then there's a trust factor with Morph. And Morph is probably in the process of building more and more trust. But if you had ZK Snarks, the inference was totally provable, then who cares? Brand has less value.

3:47

SPEAKER_2

Can you explain to the audience what ZK Snarks is? I mean, I vaguely remember this from crypto days. I wouldn't be able to explain it.

3:59

Nick Khami

You have Alice and Bob. And Alice says that she has 10 goldfish.

And Bob says, I don't know if you have 10 goldfish or not.

But Alice says, well, I can't show you that I have 10 goldfish. I don't want you to know. I don't want to make that public. So Bob gives Alice a math equation that she could only possibly solve if she had the 10 goldfish. And then Alice solves that and returns the proof to Bob. So Bob knows that Alice had the requisite information to resolve the proof. And for inference, it would be like you could only possibly have gotten the results to some equation or some formula by actually having a real deep seek.

Read the other 8 minutes of this episode here.

$15 gets you this episode and 99 more. Credits never expire. After checkout, one link brings you back here to read it.

Not what you expected? Full refund within 14 days.

Thousands of transcripts fetched by people building searchable podcast archives

Fetch transcripts with code

The same credits work from a script or an AI agent. The demo key returns a sample episode in full, no card needed:

request
curl -H "x-api-key: pt_demo" \
  https://spoken.md/transcripts/1000651996090

Markdown with the speakers named, for your notes, your knowledge base, or anything that makes HTTP calls.

From $0.10 per transcript. No subscription. Credits never expire. Prices exclude VAT, added at checkout for EU customers. Not what you expected? Email us within 14 days with 20 or fewer credits used and we refund the pack in full.

Using your own key:

request
curl -H "x-api-key: YOUR_KEY" \
  https://spoken.md/transcripts/1000801968565