CJ Muse, the Cantor Fitzgerald analyst who covers semiconductors, tells Bloomberg Surveillance that the only question that matters for the group is how long AI infrastructure spending lasts, and that on his numbers the supply side is sold out for years while the largest model builders are still short of compute. He also defends Nvidia's investments in its own customers as a supply problem rather than circular financing. This summary covers his segment of the hour; Bob Elliott and Paul Sankey, who bracketed him, are written up separately.
👤 Guest: CJ Muse, senior managing director and equity research analyst at Cantor Fitzgerald, where he covers the semiconductor sector
🎙️ Hosts: Jonathan Ferro, Lisa Abramowicz and Annmarie Hordern, who anchor Bloomberg Surveillance on Bloomberg Television weekday mornings from 6 to 9 a.m. Eastern
📰 Published: 31 August 2026 on the Bloomberg Surveillance podcast feed
🟣 Apple Podcasts | 🔗 Episode page | ⏱️ 25 min
Key Takeaways
Durability of AI infrastructure spending is the one debate that decides the whole group
"I think you're hitting on the key debate for all of semiconductors, the durability of AI infrastructure spending."
"And so our view is that this spending will continue until 2030 minimum." — and where the peak lands is what sets the multiple
The supply side is already sold out years ahead
"If you look at the supply side, TSMC is effectively sold out until 2029. Memory, similar."
"We heard from Nvidia last week that they can only meet 70% of the demand that they have for calendar 27."
The China efficiency fear is a demand-side worry, and the frontier labs are behaving as though it is wrong
"I think there's some fears around open source modeling and the effectiveness and efficiency of models in China."
"But I think internally at the largest frontier models, they see what their next gen models look like, and they are trying to get as much compute as fast as they can."
MediaTek is a supply-chain deal, not a financial one
Rick Tsai used to run TSMC and is close to Jensen Huang; "MediaTek is essentially Nvidia's physical AI partner."
Nvidia will trade away custom silicon business to get customers onto the full system rack
Nvidia's investments in OpenAI and Anthropic are a demand problem, not circular financing
"And then the other is just the simple notion that Anthropic and OpenAI have limited balance sheets." They are competing, he said, against the largest hyperscalers on Earth
"I agree more with Colette from Nvidia that this is not your typical circular financing."
The nightmare for Nvidia is a two-lab world, which is why it is going full stack
"Yeah, no, I think the worst case scenario for an Nvidia would be the world or at least the United States and non-China gravitates to a OpenAI Anthropic only frontier model world."
At that scale the labs build their own silicon, so Nvidia occupies every layer instead: "You go full stack, which means you're going to compete at every level of the AI cake that they've described."
Nvidia already looks like a hyperscaler competing with its own customers
"Nvidia is a virtual hyperscaler", in his phrase — "They are absolutely competing with the likes of Amazon, Google, et cetera."
Frontier models earn 90% of the economic profits on 30% of the tokens, and that gap invites cheaper models
"You know, I think the truth today is that the frontier models probably do about 30 percent of token generation, yet earn about 90 percent of the economic profits."
"You know, if you're just going to ask, this year, where does Thanksgiving fall? You don't need a great model for that."
Rates barely register with equity investors and register completely with the companies
"I mean, in the last few weeks, we've seen CDS spreads push out for Nvidia, Broadcom, others."
"If interest rates do continue to push higher, that will limit the magnitude of investments."
How Long the Cycle Lasts Is the Only Question That Sets the Multiple
The program introduced Muse as Cantor Fitzgerald's chip analyst and asked him to take it from the top: how durable is this cycle, given how many people doubt it and how hard he has been pushing back.
Muse agreed the doubt is the point.
"I think you're hitting on the key debate for all of semiconductors, the durability of AI infrastructure spending."
The supply side answers first, and it is sold out. "If you look at the supply side, TSMC is effectively sold out until 2029. Memory, similar."
The demand-side fear is China and open source. "And I think there's some fears around open source modeling and the effectiveness and efficiency of models in China. And is that going to reduce the overall amount of spend?"
What the largest labs do with their own roadmaps is his counter-evidence. "But I think internally at the largest frontier models, they see what their next gen models look like, and they are trying to get as much compute as fast as they can."
"Compute is sold out, right? We heard from Nvidia last week that they can only meet 70% of the demand that they have for calendar 27."
Nvidia has visibility into 2028 and 2029 as well, he said
His own call is the far end of the range. "And so our view is that this spending will continue until 2030 minimum. But that is the core debate."
The consequence is a valuation one: "But if you think this can extend into 2030 and beyond, then these stocks are incredibly cheap."
The MediaTek Investment Is Supply Chain, and the Model Investments Are a Demand Problem
A host asked how important Nvidia's funding of other companies in the ecosystem is to easing the supply issue — something the chief financial officer had raised on the earnings call — and whether shortages and bottlenecks are the right lens for reading those investments.
Muse split the answer into the MediaTek deal and everything else.
On MediaTek, the logic is relationships and physical AI. Rick Tsai used to be chief executive of TSMC and is, in Muse's words, great friends with Jensen Huang
"MediaTek is essentially Nvidia's physical AI partner. And so this combination or this investment makes perfect sense to me."
Selling the rack is worth losing some custom silicon. "I think the earlier comments around getting customers to buy into Nvidia's full system rack is, even if they have to give up some of the custom silicon business, is a strategy for Jensen."
The first of the two broader motives is locking in supply. "Number one, he's securing his supply chain."
"And I think it is multi-year contracts plus endeavoring to bring financial institutions to support the investment so that that becomes another competitive advantage for them."
The second is that his biggest customers cannot self-fund. "And then the other is just the simple notion that Anthropic and OpenAI have limited balance sheets. You know, they're competing against the largest hyperscalers on Earth. They have run out of compute."
They can see the demand for more compute to support their models, he said, so they have to find funding elsewhere, and Nvidia has been willing to provide it
He sided with Nvidia's own framing against the circularity charge. "So I agree more with Colette from Nvidia that this is not your typical circular financing. This is more of a demand problem in terms of just not enough supply. And they're helping to make that match up better."
The Worst Case for Nvidia Is a Two-Lab World, So It Is Competing at Every Layer
The show turned to open source: Nvidia bought Hugging Face shortly after its earnings announcement, and Jensen Huang has been writing letters advocating for American companies' ability to use open-source models. How important is it to Nvidia's growth that cheaper models stay available to global corporations?
Muse answered by naming the scenario Nvidia is defending against.
"Yeah, no, I think the worst case scenario for an Nvidia would be the world or at least the United States and non-China gravitates to a OpenAI Anthropic only frontier model world."
The mechanism is scale: at that concentration the labs get such a scale benefit that they build their own silicon and use less and less of Nvidia's
The defense is to occupy every layer. "So if you're Nvidia, what do you do? You go full stack, which means you're going to compete at every level of the AI cake that they've described. And as part of it, it is providing the models."
On the compute side Nvidia already competes with its own biggest customers. He said you could argue "Nvidia is a virtual hyperscaler" — "They are absolutely competing with the likes of Amazon, Google, et cetera."
On the model side the goal is CUDA adoption. If Nvidia can push its CUDA software plus smaller, enterprise and industry-focused models, he said, "that's a real win for them in terms of driving adoption of their platform."
"And really, going to first principles, that's what Jensen wants. He wants the world on Nvidia's platform."
Frontier Models Take 90% of the Profits on 30% of the Tokens
A host revived a question another guest had put the week before: is there a disconnect between the cost of building the ecosystem and the loss of pricing power at the frontier models?
Muse gave the split in economics rather than a yes or no.
"You know, I think the truth today is that the frontier models probably do about 30 percent of token generation, yet earn about 90 percent of the economic profits."
He expects that concentration to hold at the top. "You know, I think true real intelligence will garner the lion's share of the economic profits.", and as the labs keep pushing forward, "I think they'll continue to do extraordinarily well."
But most queries do not need the best model. "But there is absolutely a need for other models."
His example: "You know, if you're just going to ask, this year, where does Thanksgiving fall? You don't need a great model for that."
The routing logic he described sends each question to whichever model is best for it and offers the lowest cost per token
Financing Costs Are Invisible to Equity Investors and Central to the Companies
The last question began with the program teasing itself — "There's a near unhealthy obsession. Some people might say there is an unhealthy obsession on this program over monetary policy." — before putting higher interest rates to a chip analyst.
Muse drew a line between the two audiences.
"You know, not I guess. Not typically with equity investors, but the companies, I would think 100 percent."
The credit market has already moved. "I mean, in the last few weeks, we've seen CDS spreads push out for Nvidia, Broadcom, others."
"Cost of financing is absolutely critical for investing in these data centers."
He tied that directly to the vehicle Nvidia set up, describing it as "that $500 million platform with BlackRock, Blackstone, etc."
"Absolutely, cost of financing is incredibly important to the NPVs of these investments. If interest rates do continue to push higher, that will limit the magnitude of investments. I certainly hope they don't."
Muse's bottom line is that the semiconductor debate is not whether AI spending is real but when it peaks: compute is sold out, the largest model builders are short of it and short of balance sheet, and the multiple the market pays for the group is a bet on 2028 against 2030.
Products, Companies & Tools Mentioned
Nvidia (The center of every answer: sold out on calendar 2027 supply, buying into its own supply chain and its own customers, and going full stack to avoid a world where two labs build their own silicon)
TSMC (Effectively sold out until 2029 on his numbers, which is the supply-side half of his durability case)
MediaTek (Nvidia's physical AI partner, run by former TSMC chief executive Rick Tsai; the investment he says makes perfect sense on supply-chain logic alone)
OpenAI and Anthropic (The limited balance sheets in his argument — out of compute, competing with the largest hyperscalers, and therefore funded by Nvidia rather than by themselves)
Hugging Face (Bought by Nvidia shortly after its earnings announcement, cited by the host as evidence of how far Nvidia is backing open-source models)
Amazon and Google (The hyperscalers he says Nvidia is now absolutely competing with, which is what makes it a virtual hyperscaler)
CUDA (The software layer whose adoption is the real prize in Nvidia's model strategy)
Broadcom (Named with Nvidia as a company whose credit default swap spreads have pushed out in recent weeks)
BlackRock and Blackstone (The financial institutions in the data center financing platform he cited as evidence that cost of financing is now a competitive variable)
Cantor Fitzgerald (Muse's firm, where he covers the semiconductor sector)
Get the latest market chatter as it happens:
If this saved you the listen, send it to one person who would rather have the time back.