Athena Capital is not trying to pick the winning large language model. Isabelle Freidheim's fund buys the layer of tools that sits between the data centers and the models themselves.
The reason is pricing rather than conviction. Most of the value in the model layer has already been created and repriced in private rounds, so by the time one of those companies reaches the public market an investor is underwriting the technology itself at a price someone else set.
"We are betting on the picks and shovels. That expression comes, by the way, from the gold rush, where those that made money weren't necessarily the big miners. They were the merchants that sold the picks and shovels."
Freidheim is founder and managing partner of Athena Capital, invests in private rather than public markets, and named her most recent deal on air: Multiverse Computing, which compresses data to cut the computing power and electricity a model needs.
The full segment is covered here so you can skip it.
Here are the 4 calls that matter.
👤 Guest: Isabelle Freidheim, Founder and Managing Partner of Athena Capital, which invests in private AI companies
🎙️ Host: Paul Sweeney, who presents Bloomberg Intelligence with Scarlet Fu
📰 Published: 14 September 2026 on the Bloomberg Intelligence podcast
🟣 Apple Podcasts | 🔗 Episode page | ⏱️ length not available
Key Takeaways
Buying a large language model means underwriting the technology, and the technology is slowing
Freidheim says the risk is not only the technology but the mispricing of the companies building it
The bet she will make is on the tools needed to operate the models, not on the models
Her example is compressing data so a given result costs less computing power and less electricity
By the time an AI company reaches the public market it has already been repriced
The value was created in private rounds, which is a different investment thesis for a public buyer
A slowdown in development makes the model layer harder to invest in, not the whole sector
She still expects capital to keep flowing, into the industry forming around the models
IPO cycles run four or five years and the biggest names go first when one reopens
Smaller AI companies are looking at SPACs and direct listings as well as flotations
1. Where The Value Already Is
Paul Sweeney brought Freidheim in on the capital-markets consequence of the week's argument: whether calls from AI developers to put the brakes on would affect the deal pipeline, including the large flotations expected to come out of the sector.
Her first distinction was between the two markets. "We are particularly interested in AI investments in the private market," she said. The difficulty with the public market is that most of the value has already been created, and those companies have already been repriced privately
"So by the time they reach the public market, it's a different story. It's a different investment thesis. And you're betting on a pure transformational technology."
The opening she describes is not in the models. "When it comes to private market investments, there is a vast wealth of opportunities that exist in the middle layer. So we particularly like the layer that exists between data centers and LLMs."
2. Betting On Picks And Shovels
Sweeney asked what the Anthropic chief executive's call for a slowdown actually means for private companies. Freidheim's answer was that it makes one kind of investment harder and leaves another untouched.
"Well, that makes it harder to invest in a pure transformational technology," she said, because when you invest in an LLM "you're really underwriting the technology" — and on her reading there is now a slowdown in the pace at which that technology improves, for several reasons that are not all new
The risk she names is a pricing risk as much as a technical one. "Because of the risk that exists with underwriting a technology, the risk isn't just the technology itself. It's the mispricing of those companies."
So the position is deliberately agnostic about winners. "And the opportunity that we see is the opportunity that exists in that layer that I mentioned, where we're not betting on which LLM will win."
"We are betting on the picks and shovels. That expression comes, by the way, from the gold rush, where those that made money weren't necessarily the big miners. They were the merchants that sold the picks and shovels."
"And this is the opportunity that we underwrite, and that is that layer of tools that are needed to operate the technologies"
3. Why Capital Keeps Flowing
Asked whether AI investment continues at the pace of the past four years or so, Freidheim said probably, and gave a reason that has nothing to do with the models getting better.
"Probably, again, because there's been a lot of capital deployed in the LLM layer, but there is an entire industry that's emerging to facilitate the use of those technologies, whether it's to lower the cost, for example, or to improve the performance, to lower the computing costs, lower the electricity cost."
Her worked example is her own most recent investment. "We just invested in a company, Multiverse Computing, that, for example, will compress the data to make for lower cost of computing power that is needed and lower electricity."
"So I think, yes, the capital will continue flowing into AI. And we are now in a very different environment that we last saw in 2021, for example, when investments were being deployed into tech companies." Sweeney agreed that the two environments are not comparable
4. The IPO Window Reopening
Sweeney's last question was about timing, with Anthropic's own listing in doubt and OpenAI having said it will do nothing in 2026. "How much are these startups, these smaller companies looking at the schedule for when the big AI labs go public to determine their own timings?" he asked.
Freidheim said the smaller companies are not waiting. "They are currently looking to go public. There are a number of them that will explore IPOs. SPACs, maybe direct listings to really tap into what is a deep pool of capital that is available to them that does exist for now in the private market, but not at the size that they can find in the public market."
The motive is size: the private pool is deep, and the public one is deeper
"And right now, there's no question there is demand for these types of tech stocks. So I think companies will continue to explore."
Her framing of where the market sits in its own cycle is the closest thing in the segment to a forecast. "Again, we are just recovering from a long down IPO cycle. It's not unusual to see IPO cycles that will last four or five years. And inevitably, when the IPO cycles recover, it's very typical to have larger companies go public first."
Sweeney's response was that the recent deals were beyond large — "There were mega deals" — and that demand persists on both sides of the market
Bonus Insights
The picks-and-shovels analogy came with its own history lesson, which Freidheim volunteered rather than being asked for: in the gold rush the reliable money was made by the merchants selling equipment, not by the miners
Her definition of the middle layer is functional rather than sectoral — anything that lowers the cost of running a model, improves its performance, or reduces the computing and electricity it consumes
She placed the current moment against 2021 as the last comparable period of capital deployment into technology, and was explicit that the two are not the same environment
Freidheim's bottom line is that the slowdown argument changes what is investable rather than whether AI is investable: underwriting a model means underwriting a technology whose pace is already easing, at a price set in private rounds, while the tools that make those models cheaper to run are a bet that does not require knowing which model wins.
Products, Companies & Tools Mentioned
Athena Capital (Freidheim's firm, which invests in private AI companies in the layer between data centers and the models)
Multiverse Computing (Her most recent investment, named on air: it compresses data to cut the computing power and electricity a model needs)
OpenAI and Anthropic (The two whose listing plans set the timetable the smaller AI companies are reading — OpenAI has said nothing will happen in 2026, and Anthropic's own IPO is the one Sweeney raised as in doubt)
If this was worth your time, send it to someone closer to the industry than you are.
Get the latest market chatter as it happens: