Positron has raised $875 million at a $5 billion valuation to build inference chips that avoid Nvidia's own hardware entirely, but its chief executive says the company still wants to sit inside Nvidia's data centers, not outside them.
Rather than compete for Nvidia's customers, Positron is trying to undercut Nvidia's memory costs while staying close enough to plug into the same racks.
"Nvidia is absolutely the main player in the market, right? Whether it's training, absolutely main one, and then even in inference, they are the majority."
Mitesh Agrawal runs Positron, whose first product, Atlas, already has 50 racks deployed at Oracle.
I listened to the full segment so you can skip it.
Here are the 3 takeaways that matter.
👤 Guest: Mitesh Agrawal, chief executive of Positron, an inference-chip company that just raised $875 million at a $5 billion valuation
🎙️ Host: Ed Ludlow, who anchors Bloomberg Tech from San Francisco
📰 Published: 10 September 2026 on the Bloomberg Tech YouTube channel
🔴 YouTube | ⏱️ 8 min
Key Takeaways
Positron raised $875 million at a $5 billion valuation to scale production through a tight supply chain
Agrawal said the round, split into two tranches, funds the scale-out of its next-generation silicon and systems
Its first product, Atlas, already has 50 racks running at Oracle, sold as a straight capital purchase
The next generation, silicon called Asimov paired with systems called Titan, tapes out this year and reaches production in the second half of 2027
The company wants to sit inside Nvidia's data centers rather than replace them
Positron has an Atreides board member pushing it toward an NVLink Fusion agreement, and Agrawal says he expects to pursue one
1. Why $875 Million, Why Now
Ludlow opened on the round itself — its size, its two tranches and its investor list.
The need is capital-intensive scale in a constrained supply chain. "The need for capital is growth." Agrawal pointed to a wave of recent AI capacity announcements as evidence of how fast compute demand is rising, and said "supply chain, obviously, is a big kind of constraint," so capital goes to whichever company can scale production fastest under current conditions
The money funds the next generation, not the current one. "That's kind of what this is really driving is the scale out of our next generation silicon and systems."
2. Atlas, Asimov and Titan
The current product is already deployed. "We have our first product out recently. First generation called Atlas already out. 50 racks at Oracle getting deployed."
The next generation has two names — one for the chip, one for the system around it. "And then second generation, the silicon is called Asimov. The systems are called Titan."
The pitch is memory capacity and bandwidth, not raw compute. "What we are really, really focused on is memory-first architecture, so we drive massive amounts of memory capacity and very high realized memory bandwidth to really get fast inference, but at really good TCO."
The timeline is tight. "We are taping out end of this year with our next generation silicon, Asimov, and then Titan will be in production sometime second half of next year." First silicon typically returns "anywhere between 8 to 12 weeks" after tapeout, with TSMC as the fabrication partner
The commercial model for Atlas is simple. "Oracle buys it as a capex purchase on their balance sheet and they pay us as a PO." Agrawal said pricing evolves toward "dollar per GPU or dollar per systems per hour" and eventually "dollar per million tokens" as the company adds neocloud and hyperscale customers
Beyond Oracle, Agrawal named only categories for future customers — AI labs, neoclouds, hyperscalers, sovereigns building their own inference capacity, and companies selling or consuming tokens, video generation and inference as a service
3. Close to Nvidia, Not Against
He argued cheaper inference expands the market rather than shrinking it. Citing OpenAI's price cuts, he said: "When they slashed the prices by half, they saw the utilization go up by 10x." He said a competing model released the same week looked very cheap and strong on early results, using the same LPDDR5X memory approach Positron favors, though he had not yet examined its architecture
He does not dispute Nvidia's position. "Nvidia is absolutely the main player in the market, right? Whether it's training, absolutely main one, and then even in inference, they are the majority."
The strategy is proximity, not replacement. "For all of us, including Positron AI, we really focus on how can we work with them, right? You know, we have to be in the same data centers, in fact, even same racks as them to work with them, right?"
He expects to pursue formal Nvidia interconnect status. Asked if Positron would sign an NVLink Fusion agreement, he said: "We will definitely, like, talk to them and try it." He noted Positron's chips can already connect to Nvidia GPUs over standard Ethernet and InfiniBand without that agreement, but a closer link is the goal
A board member is pushing that view. "Gavin is on our board. He's very, very clear" that "this is a heterogeneous computer market" — Gavin Baker of Atreides
Bonus Insights
This is Agrawal's own segment of a longer Bloomberg Tech episode built around Apple's iPhone Duo launch, which also carried a separate interview with D-Matrix's chief executive on its own Nvidia inference-chip partnership
Agrawal's bottom line is that Positron is betting its memory-first architecture beats Nvidia on cost per token, without needing to leave Nvidia's own infrastructure to prove it.
Products, Companies & Tools Mentioned
Positron (Agrawal's company, maker of the Atlas chip already running at Oracle and the Asimov/Titan generation due in 2027)
Oracle (Runs 50 racks of Positron's Atlas chips today, bought as a straight capital purchase)
Nvidia (The market's dominant player in both training and inference, which Positron wants to sit alongside rather than replace)
TSMC (Positron's fabrication partner for the Asimov chip, taping out at the end of this year)
Atreides (The fund of Positron board member Gavin Baker, who Agrawal says is pushing the company toward closer Nvidia integration)
If this was worth your time, send it to someone who follows the name.
Get the latest market chatter as it happens: