enesptjazh-Hanszh-Hantfrdeitkoruarhiidtrnlplvithsvhemsukcselrohudafinobn

Reflection, Cohere and Core Automation now sell AI you own instead of rent

Canada put up to C$1.6 billion into Cohere. Misha Laskin and Jerry Tworek pitched AI customers can own, as Anthropic moves to its own compute.

By Published November 1, 2026 7 min read Grounded in the CEOInterviews.AI corpus

Over Halloween weekend three AI companies made the same pitch to three different buyers, and a fourth turned out to be acting on it. Own your AI instead of renting it. Misha Laskin of Reflection said it to American enterprises worried about Chinese open models. Aidan Gomez said it to Canadians, one day after Ottawa agreed to buy up to C$1.6 billion of Cohere. Jerry Tworek, who left OpenAI to start Core Automation, wants to sell his models by the rack instead of by the token. And according to Jay Goldberg of Seaport, the IPO filing shows Anthropic itself moving from leased computers to its own. The rental model built this industry. Those four think the next stage of spending goes to ownership, and they disagree about where ownership actually sits.

Cohere co-founder and CEO Aidan Gomez speaking on stage at Collision 2023 in Toronto
Aidan Gomez at Collision 2023 in Toronto. Photo by Piaras Ó Mídheach/Collision via Sportsfile, CC BY 2.0, via Wikimedia Commons.

Laskin's landlord analogy

Laskin went on Prof G Markets with Ed Elson, in an episode published November 1, and he explained open weights with the cleanest framing I have heard from a model company.

The high-level distinction: closed, you cannot own, you can only rent; open, you can own entirely. Right. So it's kind of like the difference between the rental and the ownership real estate market. When you are renting an apartment, you don't actually own it, right? You are subject to, you know, the whims of your landlord, which could mean price hikes.

Misha Laskin, November 1, 2026, Prof G Markets

The landlords in that picture are OpenAI and Anthropic, and Laskin was polite about them. He said they are "going to do just fine." His real target is China. Elson tried twice to get him to call Chinese open models a dumping campaign. Laskin didn't take the bait.

The fact that there are great open models out there, period, doesn't matter where they come from, is a gift to us all, right?

Misha Laskin, November 1, 2026, Prof G Markets

That sounds generous, but it is also his sales pitch. If Chinese weights are a gift, the question becomes who can use them. Laskin's answer is that big American companies can't. "It's very rare to see Chinese open models go into production in Fortune 500," he said. Startups take the risk and the incumbents don't, so Reflection wants to sell the incumbents an American model they can own. Nvidia is already one of Reflection's largest shareholders, and the Financial Times reported on October 10 that it was in talks to deepen that stake or buy the company, which was raising at a $25 billion pre-money valuation. When Elson asked about the deal, Laskin talked about the roadmap instead.

Cohere sold ownership to a government

On October 29 Prime Minister Mark Carney announced an equity investment of up to C$1.6 billion in Cohere, paid in stages. In exchange Cohere commits to at least C$1.7 billion of new computing infrastructure in Canada and at least 1,000 technical jobs by 2030. The stake is part of a Series E of up to C$7 billion, led by Schwarz Group, the German owner of Lidl, which BetaKit puts at a valuation of nearly $36 billion. Schwarz came in through Aleph Alpha, the German lab Cohere is merging with.

A day later AI minister Evan Solomon put Gomez and chief AI officer Joelle Pineau on his own podcast. The minister was not a neutral host, and he asked how this helps safety. Gomez answered that it comes down to who controls the model.

Well, I mean, the first thing to say is if you're not building it, you don't control it. You're not in a position to make the decisions about deployment, about safety, because someone else is controlling those decisions.

Aidan Gomez, October 30, 2026, Evan Solomon podcast

That is Laskin's argument, applied to a country. Pineau made it concrete. "The system is running in Canada, the data stays in Canada," she said. For a government buyer, ownership means the servers are physically in the country, under its laws.

Gomez concedes the money gap

Solomon asked the obvious question himself. How does Cohere compete with Anthropic, which he said is raising at $2 trillion? Gomez agreed that capital matters, then said the race isn't only about who spends the most.

So inside enterprises where we focus, those models need to be smaller. They actually can't be super massive because they're being deployed on that customer's hardware, and that customer doesn't have infinite hardware.

Aidan Gomez, October 30, 2026, Evan Solomon podcast

This is the honest version of the ownership argument, and it comes with a cost. Once the customer owns the hardware, the model has to fit on that hardware. Cohere is betting that banks and governments will take a smaller model they control over a frontier model they rent. I think that holds for regulated work. Whether it holds for coding agents, where people use the best model they can get, is a different question.

Tworek says owning the weights isn't enough

Tworek pushed back on the open-weight pitch, though not by name. He spent years running reasoning research at OpenAI, and he appeared in Witcher armor for a Halloween episode of Venture with Grace, published October 30. Asked how his lab differs from the open-weight crowd, he answered right away.

First and foremost we use different algorithms. The problem with open-weight models is that fine-tuning is hard, inefficient. That requires a lot of people and the results are not really guaranteed and not really very good.

Jerry Tworek, October 30, 2026, Venture with Grace

His point is that downloading weights doesn't give you a model that knows your business. Core Automation is building an architecture that is not a transformer and that learns from each customer's data while it works. He doesn't want to sell it by the token either.

And like my ideal way of charging it would be per rack. I think a lot of businesses, they like having like hardware on prem

Jerry Tworek, October 30, 2026, Venture with Grace

Each rack would carry its own memory of the customer's data, and "you know that your data doesn't leave your systems," he said. The thin part of the claim is that Core Automation has shipped no product, and he said the team still mostly codes on OpenAI models. The pitch matters anyway, because one of OpenAI's most senior researchers left to build a version of ownership he thinks the open labs can't deliver.

Anthropic is doing it too

The biggest landlord is moving toward ownership as well. On Yahoo Finance on October 31, Goldberg explained why he wrote the first analyst note on Anthropic but gave it no rating. He pointed at a warning in the S-1.

They've actually explicitly warned that the model is going to shift, and they will be profitable probably in the next quarter or two, but then they're going to start going back into losses, because they're changing the way that they approach acquiring computing resources.

Jay Goldberg, October 31, 2026, Yahoo Finance

Anthropic used to lease compute by the hour from Amazon and Google, Goldberg said, and it now treats owning compute as a competitive weapon. He thinks the larger risk falls on the clouds. If enterprises start choosing a cloud by "which one's going to get me the best deal on Anthropic," the cloud becomes the reseller.

And Anthropic has done a great job of going through the partners, but increasingly I think direct is going to be preferred.

Jay Goldberg, October 31, 2026, Yahoo Finance
Reflection co-founder and CEO Misha Laskin on Prof G Markets with Ed Elson
Misha Laskin on the Prof G Markets Founder Series, November 1, 2026. Thumbnail via YouTube.

Who ends up owning what?

Each of them puts ownership at a different layer. Laskin says it is the weights. Gomez and Pineau say it is the servers on Canadian soil. Tworek says it is the rack and the memory inside it. Anthropic's version is the data centers under its models, and that move turns its cloud partners into competitors. What they have in common is a buyer who no longer wants to pay a markup to someone else's landlord.

My read is that the ownership pitch wins wherever the buyer has a regulator or a flag: banks, defense and governments, like the one that just bought into Cohere. It loses wherever the frontier model is clearly better, because nobody runs a weaker coding agent out of principle. Watch who buys Reflection's models in production next year. Laskin says that isn't happening yet with Chinese weights in the Fortune 500, and nobody has shown yet that it happens with American ones.