Conversation
Highlights

How Positron AI Sells Silicon Before Shipping: Lessons from Competing Where Nvidia Owns 90%+

The traditional silicon playbook says: design for three years, tape out, manufacture, then sell. By the time you reach customers, your architecture is already dated.

Mitesh Agrawal is inverting this sequence. In a recent episode of BUILDERS, the CEO of Positron AI explained how his company is securing purchase commitments from anchor customers before their chip even ships—and why pre-sales create organizational momentum that extends far beyond revenue validation.

The Memory Bandwidth Arbitrage

When one player controls 90%+ of a market, their optimization decisions create exploitable gaps. Nvidia’s challenge isn’t technical capability—it’s revenue concentration.

Training workloads still drive over 60% of silicon revenue. These workloads are FLOP-bound, so Nvidia’s roadmap has prioritized compute density over memory bandwidth for the past decade. “Over the past decade the amount of flops on the chip have grown at a much faster rate than the memory bandwidth and memory capacity on the chip,” Mitesh explained.

But inference workloads—particularly decode-heavy applications generating long outputs like reasoning models, code generation, and video synthesis—hit a different bottleneck. Every token generation requires loading model weights from memory. When OpenAI’s CFO noted their compute grew 10x from 0.2 to 2 gigawatts while revenue scaled identically from $2B to $20B, she was highlighting how inference economics work: “It’s almost like a linear increase. It’s no coincidence that’s the case.”

Positron’s technical bet is straightforward: memory bandwidth and capacity matter more than FLOPS for these workloads. Their upcoming chip delivers 2TB of on-chip memory versus Nvidia Rubin’s 0.4TB. “We picked one domain and we’re not, I’m not sitting here, I’m going to be grounded. We’re not going to be outperforming Nvidia across all workloads,” Mitesh said. “But in that one domain we are trying to really expand the Pareto curve by 3, 4, 5x performance per dollar and performance per watt.”

The lesson for founders: dominant players optimize for existing margin dollars, not theoretical future workloads. Map technical bottlenecks against their incentive structure to find wedges.

Your Best Positioning Comes Post-Launch

Positron taped out their first generation in 15 months—matching Nvidia’s 12-15 month cadence rather than the industry’s 3-4 year norm. They pitched performance-per-watt efficiency. Customers nodded politely.

Then came deployment. Customers revealed what actually drove their decisions: air cooling capability.

“We launched our product, it’s very energy efficient, so we’re able to air cool it,” Mitesh said. “A lot of the existing state of the art systems are all liquid cooled, whether that’s Nvidia, Blackwell, Rubin, AMD’s generations and things like those.”

The implications compound. Liquid cooling requires infrastructure rebuild—new data center designs, cooling systems, power delivery. Air cooling monetizes existing capacity immediately. “If you had looked at our pitch deck like a year ago when we launched our first product, is nowhere in the conversation other than saying performance for watt efficiency,” Mitesh admitted. “Once we learned that, it became a very key messaging for us.”

This wasn’t about features—it was about removing capital deployment friction. Customers weren’t optimizing for peak performance; they were solving for “how do I utilize assets I already own?”

Design systematic post-deployment debriefs. Your second product’s positioning should reflect customer language about your first, not your original thesis.

The Organizational Physics of Hardware Pre-Sales

Positron’s second-generation product launches end of 2025 or early 2026. Mitesh’s top priority isn’t product completion—it’s securing anchor customers willing to commit purchase orders now.

“The biggest aim for the company is how can we drive sales ahead of product,” he explained. “Can we find one or two anchor customers that are like yeah, this is a great architecture, great technology, you convinced me on the technology of it. We are willing to actually take the execution risk from your perspective.”

This isn’t just revenue validation. Pre-sales create second-order effects throughout the organization.

“If you have someone to build for, the fillip it gives the engineering team, the amount of confidence it gives our operations and our supply chain vendors to really grow that production out,” Mitesh said. “Sometimes we underwrite that kind of that fill up that you get.”

Engineering teams build against specific customer requirements rather than theoretical specifications. Supply chain partners commit manufacturing capacity based on proven demand. The entire operation aligns around deployment timelines, not estimated TAM.

“It takes a lot of guts on both the seller and the buyer side of things to do it,” Mitesh acknowledged. “But the rewards are also high. It’s kind of like venture capital investing. You’re kind of putting money in a company that you think has great idea and you think have great people.”

The buyer is taking execution risk with you. That commitment signals conviction internally. Use it.

Selling Means Co-Creating Internal Business Cases

Technical proof doesn’t close hardware deals. Internal champion success does.

“It’s all about storytelling actually ultimately,” Mitesh said. “You teach them and you tell them about your technology, you tell them why you think it’s going to be successful, you show them demos, you show them emulations, you show them white papers, you show them all the reasons why it will be successful, okay, great. But then it becomes like, well, if I sell it to you, why will it be useful to you?”

This is where months of work happens. For cloud providers: what workload mix does this enable that attracts new customer segments? For enterprises: which cost centers does this eliminate? The answers require building customer-specific TCO models, infrastructure deployment scenarios, and competitive positioning frameworks.

“Building that kind of story together, that memo together, that presentation together, that just talking to people there to, you know, every waking conversation with them, it’s like, hey, this is what’s going to drive your next generation of growth for you,” Mitesh explained.

Your champion needs ammunition for budget committees, technical validation for architecture reviews, and ROI models for CFO approval. The sale happens when they can confidently present your solution through their org’s approval process.

Invest in making your champion successful internally, not just in convincing them personally.

Stay Engineering-Heavy Until Demand Is Proven

Positron has 50 people. Exactly 47 focus on engineering. “We don’t have product people or we don’t have a lot of sales. We have one salesperson per se which is an external facing person,” Mitesh said.

This structure is deliberate. “It was a very conscious decision to do it and that’s kind of something that we will do for the long run.”

The CEO handles most customer conversations. The CTO presents at conferences. “Almost everyone has to sell, including our CTO,” Mitesh noted. When their technical team speaks publicly, they’re demonstrating architecture and selling value simultaneously.

Most deep-tech founders face pressure to build “balanced” organizations. Resist this until founder-led selling proves repeatable demand patterns. Premature sales hiring means paying people to learn your market while product-market fit remains uncertain.

Let customer-facing roles emerge from specific, proven needs rather than theoretical org charts.

Markets Want Oligopolies Even When One Player Is Superior

One insight shapes Positron’s entire competitive strategy: “Market doesn’t want to default to monopoly unless that monopoly is willingly out of the like. Markets actually want to find, you know, other vendors other things. They want oligopoly structure to exist.”

Customers seek alternatives even when Nvidia delivers superior performance. Not because they expect you to win across all benchmarks, but because they need supply chain resilience, pricing leverage, and infrastructure optionality.

“We know what Nvidia is doing, we know what Google TP’s are doing, we know what other silicons are doing. And look, they’re especially Nvidia is absolutely the best silicon in the world for all types of applications,” Mitesh acknowledged.

The strategy isn’t claiming broad superiority. It’s establishing credible excellence in specific domains while the incumbent optimizes for broader, more profitable segments.

“Beyond Nvidia and AMD, no company in the AI silicon world has been able to successfully sell to an Oracle, to a Microsoft, to Google and Amazon,” Mitesh said. “It’s not trivial job. But if we want to be successful, that’s where you find really that really strong kind of adhesion from which you can grow.”

Position as the credible second option for specific use cases. Customers will create room for you if you solve real constraints.

Measure Success in Deployment Volume, Not Valuation

When asked about ten-year vision, Mitesh focused on one metric: deployed chips.

“To me it’s like how many chips can we get out? Nvidia is over 10 million, 15 million chips. If you can get anywhere close to that in the next 10 years, that’s a goal worthy of doing,” he said. “Whether we will see what the enterprise value will see, what the revenue that leads to. All those things are secondary.”

The path requires relentless product iteration. “If you tell me that in 10 years time Mitesh, I guarantee you I’ve seen the future… in the next 10 years you have launched over 10, 12 products in the silicon world, I will give much more probability that we will be successful.”

This matches Nvidia’s cadence and defies the traditional silicon industry’s 3-4 year cycles. It requires architecting your entire development process—design tools, verification flows, supply chain relationships—for annual iteration rather than perfect individual releases.

“You have to constantly do every 12 to 15 months,” Mitesh emphasized. Speed of learning compounds more than perfection of individual products.

Companies don’t win concentrated markets through single perfect products. They win through continuous learning cycles, rapid adaptation, and compounding advantages from shipping frequently.

Positron AI’s approach reveals how to compete against dominant incumbents: identify architectural gaps created by their economic incentives, discover your real value proposition through customer deployment, secure pre-sales to create organizational momentum, stay engineering-focused until demand patterns prove repeatable, and architect for rapid iteration over individual product perfection. In markets where one player controls 90%+, execution speed and domain focus create more leverage than resource advantages.