Dylan Patel on AI Labs, Chip Competition and the Race for Superintelligence

Matthew Berman speaks with Dylan Patel, founder and CEO of SemiAnalysis, about AI labs, chips and the business of computing. The conversation was published on June 30, 2025; Patel’s descriptions of private decisions and future outcomes are his reporting, interpretation or speculation.

Berman asks why Meta’s Llama 4 launch did not meet some expectations and why Behemoth was delayed. Patel distinguishes the three announced models and says he is uncertain whether Behemoth will ever be released. His assessments of their training and comparative quality are his own, not a disclosed Meta diagnosis.

Patel alleges that one model was rushed in response to DeepSeek and that its mixture-of-experts routing left some experts little used. He treats that as evidence of a training problem. The interview does not independently verify the internal cause or identify an official explanation.

Berman asks how a company with strong researchers and abundant compute could struggle. Patel says organization is hard: a team must choose which ideas to pursue and give technically qualified leaders responsibility for evaluating alternatives.

Patel points to OpenAI’s technical leaders as an example of people selecting directions from competing proposals. Individual researchers naturally favor their own ideas; a lab still needs someone to decide what deserves resources and what should be abandoned.

Once a group commits to an unproductive direction, Patel says, teams may keep exploring its branches rather than reversing course. He calls the ability to judge promising work “taste”: small-scale results do not necessarily carry over when the run expands from hundreds to many thousands of GPUs.

Patel compares choosing research directions to deciding which critic to trust. Great individual researchers can still produce disappointing results if decision-making, incentives or internal politics put the wrong choices in charge.

Berman turns to Meta’s Scale AI transaction and recruiting reports. Scale confirmed a significant Meta investment and founder Alexandr Wang’s move to Meta; contemporary reports described a minority stake, not a complete acquisition. Patel believes recruiting leadership was a major motive, but that interpretation is his own.

Patel says he has heard some other AI labs were reconsidering Scale’s services because of competitive data concerns. His specific customer-spending estimates and claims about private communications are not established by the interview.

Patel argues that Wang brings organizational and product skills and that Meta’s apparent priorities have shifted toward superintelligence. Berman asks why that phrase has gained ground over AGI. Patel says definitions of AGI have become inconsistent and sees the terminology as partly a rebranding.

They discuss reports that Meta approached other AI companies and people. Patel explicitly presents the possible negotiations, decisions and personal motivations as rumors or guesses; they are not confirmed transactions in this conversation.

Patel suggests recruits may value both compensation and the ability to shape AI products at a large company. He characterizes Wang and former GitHub CEO Nat Friedman as strong product and organizational leaders rather than judging them solely as model researchers.

Berman asks whether very large compensation offers can build a strong research culture. Patel says acquiring a small research organization also effectively assigns a very high value to its people. He disputes the claim that no leading researchers would move, and repeats an unverified figure he says he heard about an individual offer.

Patel frames Meta’s spending as a strategic choice if its leaders believe advanced AI will be economically important. The scale of individual offers, the identities of recipients and any eventual success remain uncertain.

Berman asks about the OpenAI–Microsoft relationship. Patel describes a complex partnership involving investment, access to intellectual property, revenue or profit arrangements and compute. He repeatedly says he is recalling details from memory; the precise percentages and caps he mentions should not be treated as verified contract terms.

Patel says Microsoft’s former exclusivity on compute was relaxed and that OpenAI could pursue additional infrastructure. Public announcements in January 2025 described a right of first refusal and the Stargate project. The companies later renegotiated their partnership, so the discussion is a snapshot of that period.

Asked what Microsoft received in exchange, Patel points to the reported right of first refusal and speculates that competition and capacity considerations mattered. He says OpenAI wanted more compute than Microsoft could supply at the desired speed, an account he does not independently document here.

Patel sees intellectual-property access and the definition of AGI as potential sources of tension. He sketches hypothetical outcomes in which Microsoft retains valuable access even as OpenAI seeks more investment; these are scenarios, not determinations of contractual rights or company value.

He says OpenAI’s need for capital makes a complex corporate and financial structure harder to explain to investors. His revenue and profitability projections are forward-looking estimates, not established future results.

Berman asks about GPT-4.5. The speakers initially disagree about whether it has already been retired. OpenAI had announced API deprecation in April 2025, citing GPT-4.1’s comparable or improved performance on many capabilities at lower cost and latency; that was distinct from availability in ChatGPT.

Patel says the model, known as Orion, emphasized large-scale pretraining. He finds it good at some natural conversation but too slow and expensive for many uses relative to alternatives. That is his assessment of its utility, not an independent measure of all tasks.

He argues that a large parameter count without sufficient new data can lead to memorization and eventually diminishing returns. He presents this as an explanation for GPT-4.5 and claims early checkpoints performed strongly before progress slowed; the interview does not verify the private training details.

Patel also alleges a training-code bug and repeated restarts from checkpoints on a complex run. He says coordinating a large infrastructure fleet is difficult even before considering model design. These are his assertions about internal events, not confirmed disclosures in the interview.

He invokes the Chinchilla scaling work as an example of balancing model size and training data for a compute budget. Patel notes that an often-quoted tokens-per-parameter ratio was tied to specific architectures and is not a universal rule for all modern models.

Patel contrasts pretraining with later reasoning-oriented approaches. He argues that generating candidate solutions in domains where answers can be checked can provide useful training signals. His account compresses several different training and inference techniques and is an interpretation, not a complete explanation of any company’s model.

Berman asks why Apple appears behind in frontier models. Patel describes Apple as historically cautious about large acquisitions and says its secretive culture can make recruiting researchers who want to publish harder. These are his judgments about organization and talent, rather than an audited assessment of Apple’s internal work.

Patel says established research communities attract further talent, making it difficult for a new entrant to assemble a top team. He compares Apple’s situation with the research cultures at Google DeepMind, OpenAI, Anthropic and Meta.

Patel recounts historical tension between Apple and NVIDIA, including defective laptop GPUs and patent disputes. He says he does not remember every detail and does not assign fault for the hardware failures. His inference that these events explain present company strategy is speculative.

Berman asks about on-device versus cloud AI, citing privacy and latency as advantages of local processing. Patel is skeptical of the economics for the most demanding models: local hardware is constrained by memory bandwidth and cost, while many cloud products offer free tiers.

Patel says a consumer may prefer paying for a phone with less specialized hardware and using a cloud model. He argues that many valuable tasks, such as searching maps, checking a calendar and booking travel, already require network access and data from remote services.

For short interactions such as keyboard prediction, Patel accepts that local inference may be useful. For long research or booking tasks, he expects the strongest model and the external tools to run in the cloud; his generalization about consumer preferences is a forecast, not evidence that privacy never matters.

Berman sees a case for a mix of local and cloud processing. Patel agrees some work will run on devices, especially small recognition and tracking tasks for wearables, but expects the most valuable reasoning to skew toward cloud infrastructure.

Patel discusses Apple’s cloud infrastructure and says the company will also need remote processing even while promoting on-device features. His specific account of staffing and future hardware plans is presented as his industry analysis, not an Apple announcement.

The discussion turns to NVIDIA and AMD. Patel says AMD has strengths but trails NVIDIA on developer experience and system integration, though he expects it to gain some share. His conclusions depend on hardware generation and workload and should not be read as timeless rankings.

Patel emphasizes the interconnect: NVIDIA’s closely connected systems can combine many GPUs, while AMD systems he discusses have smaller tightly connected groups. He says this affects distributed training and inference as much as an individual accelerator’s specifications.

NVIDIA’s advantage, in Patel’s view, extends beyond CUDA itself. A user may invoke an inference library such as vLLM or SGLang, which relies on deeper layers including PyTorch, compilers and hardware-specific kernels. The quality of those layers shapes how easy it is to deploy a model.

Patel says AMD software is improving but can require more tuning choices to reach good performance. He notes that large labs can invest in low-level engineering, whereas many customers want to supply model weights and run them with minimal configuration.

Patel describes NVIDIA’s relationships with major cloud providers and smaller “neoclouds.” In his account, broader GPU distribution can create competition in cloud pricing. His quoted hourly costs and margins are illustrative estimates, not a verified current price comparison.

He says NVIDIA’s Lepton software and DGX Lepton marketplace could put it in tension with cloud companies by helping rent spare GPU capacity. Patel reports dissatisfaction among some providers and speculates about NVIDIA’s future deployment; those private reactions and plans are not independently established here.

Patel also describes arrangements in which AMD sells accelerators to cloud providers and rents capacity back. He says such transactions can help establish demand and customer confidence, and explicitly distinguishes the business arrangement from an allegation of illegal accounting.

He expects AMD to sell substantial volumes without rapidly overtaking NVIDIA. For a buyer, he says the answer depends on price, workload and the engineering available to optimize software. He cites Meta as an example of a company using both vendors.

Berman asks about xAI and Musk’s claims for an upcoming Grok model. Patel says he has heard it is good but does not know its final performance. He was pleasantly surprised by Grok 3 and sometimes uses its fast research features; statements about a future version remain speculation.

Patel says he likes exploring economic and historical questions and sometimes finds Grok more willing to discuss subjects he wants to investigate. His examples are personal impressions of model behavior, not evidence that any model’s answers or historical explanations are necessarily accurate.

For everyday queries Patel says he often uses OpenAI o3 or Claude 4, and at work uses Gemini for long documents, permits and regulatory filings. His choice depends on whether he prioritizes speed, reasoning or document context.

He says xAI has accumulated substantial compute and invested aggressively in data-center capacity and power, but declines to predict whether the next model will lead or merely be competitive. Berman asks whether its approach differs fundamentally from other labs.

Patel thinks major labs generally use large transformer pretraining and reinforcement learning, often in verifiable domains such as code and mathematics, while exploring computer use and less easily verified tasks. He treats differences as meaningful research choices rather than claiming identical implementations.

On employment, Berman raises forecasts that many white-collar jobs could disappear. Patel says productivity gains could ultimately reduce working hours and improve living standards, but distribution of the benefits is a central challenge. His broad historical comparisons are context and opinion, not a forecast of equal outcomes.

Patel is excited about robotics because difficult physical jobs remain hard to automate, while some creative work already faces generative tools. Berman notes that useful robotics may still take time to deploy. Neither provides a verified schedule for particular occupations.

Patel expects AI tasks to grow longer, eventually lasting hours or days, and speculates that some workflows may ultimately require no human in the loop. He is cautious about timelines and suggests 20% job automation might occur around the end of the decade or early in the next; this is a prediction, not an established outcome.

Berman questions how quickly capability turns into deployment and what happens to junior software engineers. Patel sees rapid adoption of coding assistants and worries that junior hiring may shrink even when companies invest productivity gains in more work.

Patel describes SemiAnalysis growing while using AI to increase output. He contrasts experienced engineers directing tools with the challenge of giving junior staff enough guidance, while recognizing that younger workers may adapt quickly. This anecdote does not establish labor-market causation or a universal hiring pattern.

He encourages junior developers to build skills and demonstrate work with new tools. Berman points out that not everyone wants or can afford to become a self-directed builder, and Patel acknowledges that some people need a job and initial guidance.

Patel predicts that closed models may ultimately dominate and speculates that Chinese developers could publish fewer weights if they take the lead. He hopes the market does not concentrate into only a few companies. These are forecasts and concerns, not settled facts.

Asked to pick the first company to reach superintelligence, Patel chooses OpenAI, then Anthropic, with Google, xAI and Meta also contenders. He thinks additional breakthroughs beyond existing reasoning methods will be needed. The ranking and the concept’s timeline are speculative.

Patel suggests Anthropic has become less conservative about releases and says Meta’s hiring could make it more competitive. Berman thanks him for the discussion.