SemiAnalysis: Anthropic Is Already Profitable, Memory Shortage Could Run for Years, and Nvidia's Blackwell Beat Its Own Hype by 30x
Dylan Patel of SemiAnalysis joins the Next Big Thing podcast, July 2026, to map bottlenecks across compute, memory, networking and power
Dylan Patel, founder of SemiAnalysis, sat down with WisdomTree's Next Big Thing podcast for a wide-ranging conversation that doubled as a state-of-the-union for AI infrastructure investors. Patel's firm has grown from a two-person newsletter in 2022 to a 90-person research shop with over $50 million of donated hardware from Nvidia, Microsoft, Amazon, Google, Oracle, CoreWeave, Nebius and Crusoe, and its benchmarking data has become influential enough that Jensen Huang cited it directly on stage at GTC. The conversation ranged from the economics of frontier labs to the mechanics of memory pricing, CPU demand, optics and power, and it produced several data points that move the needle for investors positioned across the AI supply chain.
Anthropic Has Turned the Corner on Profitability
The most concrete new data point in the conversation concerns Anthropic. Patel said the company was free cash flow positive and profitable in April and May of 2026, with June "looking like it's going to be the same way" even though the books were not yet fully closed. Annualized recurring revenue has "soared past $50 billion," and gross margins are "above 70%." That is a meaningfully more constructive picture than the market's default assumption that frontier labs are structurally cash-incinerating. Patel was careful to note this is not universal across the industry, but pointed to OpenAI's revenue also inflecting as adoption of its Codex coding product and other agentic tools accelerates.
Nvidia's Blackwell Beat Its Own 25x Claim
Patel recounted a moment that went viral within the industry: at GTC in March, Jensen Huang held up a SemiAnalysis-branded "inference king" belt on stage and spent roughly five minutes discussing SemiAnalysis's independent benchmarking of Blackwell versus Hopper. When Blackwell launched, Huang claimed a 25x performance improvement, a number the market and even Patel himself dismissed as marketing. SemiAnalysis's InferenceMAX benchmarking suite, run nightly across donated hardware from every major cloud, eventually showed Blackwell running DeepSeek V3 workloads 30x faster than Hopper. "Jensen, I was wrong. You were sandbagging it. It was 30x," Patel told him. The episode is more than color: it is independent, reproducible confirmation that Nvidia's generational performance claims, often dismissed by skeptics, have proven conservative rather than inflated.
SemiAnalysis's Own AI Spend Is a Live ROI Case Study
Patel offered his own firm as a real-time data point on the AI spending debate that has unsettled investors after recent commentary questioning enterprise AI ROI. SemiAnalysis's internal AI tooling spend, what he calls "annual recurring spend," was under $100,000 in November 2025, when the firm was largely just paying for standard chat subscriptions. Once Claude Code hit its stride with Opus 4.5 and 4.6, that run rate jumped to $4 million by the end of January 2026, and now sits near $11 million, with peak weeks annualizing to $14 million, for a firm of 90 people. "That's freaking insane, right?" Patel said, noting AI spend is already roughly a third of employee compensation costs and could approach half by year-end. He argued the ROI is real at his firm because it has translated into shipped product and revenue growth, but acknowledged many companies have "blown through" their entire annual AI budget by mid-year and are now choosing between cutting AI spend, cutting other software licenses, or cutting headcount instead. Companies that clamp down on AI spend, he warned, "are going to get left in the dust in terms of productivity gains."
Why Token Efficiency, Not Price, Is Deciding the Model War
Patel offered an explanation for why Anthropic continues to win enterprise coding workloads despite OpenAI's models sometimes having an edge on raw benchmarks. The key variable is token efficiency, not headline capability. Where OpenAI's models can occasionally out-execute Anthropic on frontier science, math or coding tasks, they typically take three times as long and four times as many tokens to do it, which raises cost and slows the human-AI feedback loop. Anthropic's more token-efficient models, he said, are why SemiAnalysis "remains a majority Anthropic shop," reserving OpenAI's Codex mainly for tasks that can run unsupervised overnight. He also noted that, counterintuitively, cost optimization for AI-assistant workloads often means adopting the newest model rather than a cheaper one, since a more capable model can finish a task in a quarter of the tokens and a single exchange rather than several rounds of back-and-forth.
Memory's Shortage Is Structural, Not Cyclical
On memory, Patel pointed back to a January 2026 SemiAnalysis note that argued the market was underestimating the durability of the current upcycle. His framework: memory capacity is only growing 20% to 30% a year over the next three years, while AI demand for memory is doubling. That imbalance, he said, forces less price-elastic buyers, smartphones and laptops among them, out of the market to make room for AI. Chinese smartphone makers such as Xiaomi have already seen shipments in the mid-range and low end fall roughly 40%, but the high end has not yet been squeezed, meaning iPhone and MacBook prices will need to rise by "a few hundred dollars," not just $100, before equilibrium is reached. Patel was direct about where this leaves valuations: "Memory isn't a shortage and this is not a short-term shortage. It's a shortage that's going to last years." He also flagged that memory gross margins, not yet at the 85% to 90% level he expects them to reach, will eventually mean-revert, but not before the current upswing runs further.
CPUs: Real Demand, But a Catch-Up Trade, Not a New Paradigm
SemiAnalysis was early in flagging a shift in CPU demand tied to reinforcement learning and agentic workflows, which require far more CPU-side compute for environment checking, tool calls and code execution than earlier chat-style inference. That call, made in institutional research in November 2025 and in the public newsletter in January 2026, has coincided with sharp rallies in Arm, Intel and AMD. But Patel pushed back on where sell-side enthusiasm has taken the narrative, saying that some analysts have wrongly concluded CPU spend is becoming comparable to GPU spend. "The sell side, who doesn't really understand technology at all, is just making up stuff," he said, noting that even at an improved ratio, CPU spend remains a small fraction of GPU dollars, roughly $5,000 per CPU against $50,000 per GPU in his illustrative math. What is really happening, he argued, is a one-time catch-up: hyperscalers under-bought CPUs relative to the GPUs they deployed in 2023 and 2024, and are now correcting that backlog, a dynamic he expects to normalize into a steadier growth rate once the catch-up is complete.
Networking: Copper Buys Time as Co-Packaged Optics Slips to 2029
Patel used a SemiAnalysis institutional note published earlier in the week to argue that the market has become too optimistic on the timing of co-packaged optics, a technology many investors have treated as an imminent, sweeping upgrade cycle. "Currently people are a little bit too excited on CPO," he said. "It's not coming in '27 in my view. Really in the tail end of '28, but '29 is the real ramp for scale-up co-package optics." Manufacturing yields, chip readiness and volume all remain immature, and Nvidia's own roadmap reflects the delay: Rubin and Rubin Ultra remain copper-based, with Feynman, the generation after, still not fully committed to CPO on the GPU. The near-term beneficiaries of that delay, per Patel, are copper interconnect suppliers such as Amphenol, which SemiAnalysis now expects to outperform prior estimates, alongside conventional (non-CPO) optical transceiver makers, even as the long-term destination for the industry remains co-packaged optics.
Power: An Improvisation Era for Data Center Energy
On energy, Patel said data center capacity additions are set to roughly double again, from 20 gigawatts this year to 30 gigawatts next year and 50 gigawatts the year after, and that within roughly two years, half of the incremental power for new data centers will be generated on-site rather than drawn from the grid. Transmission remains the hardest bottleneck to solve given regulatory and utility-amortization constraints, but generation and power conversion are seeing what he described as a wave of improvisation: repurposed diesel truck, train and marine engines converted to run on gas and back-driven into electrical generators, reciprocating engines and industrial gas turbines alongside traditional combined-cycle gas plants from GE Vernova, Mitsubishi and Siemens. Solar-plus-battery costs are falling fast enough, aided by China's manufacturing scale, that Patel expects the technology to undercut gas within about two years, depending on the reliability threshold operators demand. On the more exotic end, SemiAnalysis has also published research on data centers in space, which Patel noted removes the battery requirement entirely. "It's going to be a pain in the ass, but it will work," he said of the diesel-conversion approach, a fair summary of his broader view that power, not chips, is now the more improvisational constraint in the buildout.