The claim travelling around the AI industry is dramatic: NVIDIA's CUDA moat is “rapidly crumbling.”
The argument behind it is more interesting than the headline.
It comes from a leaked transcript of a nearly four-hour conversation between DeepSeek founder Liang Wenfeng and investors, reportedly held on May 20 and circulated publicly on July 23. The transcript says explicitly that the text has not been confirmed by Liang. It appears to have passed through transcription and translation, so the sensible approach is to analyse the ideas without treating every sentence as an authenticated quotation.
Read that way, the conversation reveals a coherent DeepSeek strategy built around four convictions:
- AGI, not short-term revenue, is the destination.
- Continuous learning is the main capability missing from today's agents.
- Open weights and low API prices can be a durable business model.
- AI-assisted software and new compiler layers will make accelerator ecosystems easier to reproduce.
Liang is directionally right on several of these points. But his own comments also contain the best case against the most aggressive interpretation of them.
What Liang Actually Said About AGI
Liang's definition of the current limitation is unusually practical.
Today's best models can outperform people on many bounded tasks when they receive complete context and instructions. The problem is that real work rarely arrives in that form. A new employee learns who matters, how the organization behaves, what happened last quarter, which exceptions are normal, and when not to follow a process literally. A model is usually asked to reconstruct that accumulated context for every task.
Liang describes AI progress as a ladder:
language models, then chain-of-thought reasoning, then agents, then continuous learning.
Agents extend what models can do by giving them tools, time, and the ability to act. But they still plateau if they cannot retain and learn from experience over long periods. For Liang, AGI begins when a system can learn continuously, perform the range of work a person can perform, and eventually contribute to the development of its successor. Embodied intelligence comes later.
This framing deserves attention. It moves the AGI discussion away from a single benchmark or an arbitrary intelligence threshold and toward organizational usefulness. A capable employee is not valuable only because of what they knew on day one. They become valuable because experience compounds.
The difficulty is that continuous learning is not simply “more memory.”
Western products are already moving toward longer contexts, persistent memory, tool use, and multi-day agentic work. Those mechanisms let a model retrieve previous experience without changing its underlying weights after every interaction. True online learning is harder. A production system must decide which experiences are trustworthy, prevent data poisoning, avoid catastrophic forgetting, preserve safety behaviour, respect deletion requests, and make learned changes auditable.
Liang has identified a real frontier. He has not shown that it is one clean step away.
The CUDA Argument Is About Abstraction
Liang's CUDA thesis has three parts.
First, AI can write and translate more of the low-level software needed to support a new accelerator. That lowers the labour cost of recreating kernels, libraries, and integration code.
Second, higher-level systems such as TileLang allow developers to describe the data flow of an AI kernel while compilers handle more of the hardware-specific scheduling. The published TileLang work is explicitly designed to separate kernel data flow from low-level scheduling decisions. In principle, that makes the same model architecture easier to move across NVIDIA, Huawei, AMD, or future specialist chips.
Third, AI accelerators are becoming purpose-built systems rather than adaptations of graphics processors. If future chips are designed around transformer, mixture-of-experts, and long-context workloads from the beginning, their software layers do not need to inherit every CUDA-era assumption.
All three forces are real.
AWS is already adding AI-assisted model-porting and equivalence-checking tools around Trainium. Google's seventh-generation Ironwood TPU can run PyTorch and JAX in pods of up to 9,216 chips. AMD continues to expand ROCm, including HIP as a portable language for AMD and NVIDIA GPUs. The Western market is not waiting for CUDA to disappear before it develops alternatives.
But easier kernels are not the same thing as a complete production ecosystem.
NVIDIA's Moat Has Moved Up the Stack
CUDA began as a programming model. NVIDIA's current advantage is a system.
It includes optimized libraries, profilers, debuggers, model integrations, deployment containers, enterprise support, and a large population of engineers who know how the stack behaves under pressure. For frontier workloads, it also includes the physical system around the chips: NVLink, networking, memory movement, distributed scheduling, failure handling, and rack-scale operations.
TensorRT-LLM now combines custom kernels with quantization, paged caches, speculative decoding, expert parallelism, and multi-node serving. NVIDIA Dynamo adds request routing, disaggregated prefill and decode, distributed cache management, and resource scheduling. Rubin combines GPUs, CPUs, switches, network interfaces, storage, and software as one co-designed platform.
Replacing a CUDA kernel may become much easier.
Replacing the economics and reliability of that whole system is a different project.
This is why “crumbling” is the wrong time horizon. The moat is becoming more permeable at the programming layer while NVIDIA is rebuilding it at the system layer.
Liang effectively concedes this. In the same discussion, he reportedly says DeepSeek's domestic hardware remains about two years behind, that several Huawei accelerators are needed to replace one comparable NVIDIA unit, and that available capacity is sufficient for current work but not its next-generation model. He also says the TileLang-based transition is not finished.
That is not a collapsed moat. It is a credible migration programme under severe constraint.
China and the US Face Different Economics
The most important context is not technical. It is geopolitical.
Liang says plainly that domestic substitution would be difficult in a normal commercial environment where NVIDIA hardware was readily available. Restricted access changes the calculation. Chinese labs must absorb the cost of immature software, lower chip performance, and complex porting because strategic autonomy is worth more than short-term efficiency.
US labs and cloud providers have another option: diversify without leaving NVIDIA.
Google can train and serve models on its own TPUs while also offering NVIDIA infrastructure. AWS can develop Trainium and continue deploying NVIDIA systems. Microsoft, Oracle, CoreWeave, and other providers can introduce custom accelerators for selected workloads while buying Rubin at scale. AMD can win workloads through ROCm without requiring customers to abandon every familiar framework.
The Western pattern is therefore likely to be heterogeneous rather than post-NVIDIA.
Custom silicon will take the workloads where volume, predictability, and control justify specialization. Portable compiler layers will reduce switching costs. NVIDIA will remain strongest where frontier performance, rapid model change, broad compatibility, and operational maturity matter most.
China is trying to cross the river because it must.
The US is building more bridges while keeping the fastest existing one open.
DeepSeek V4 Versus the Current Frontier
DeepSeek's latest public family is V4, released in April. V4-Flash has 284 billion total parameters with 13 billion active; V4-Pro has 1.6 trillion total with 49 billion active. Both offer a one-million-token context, up to 384,000 output tokens through the API, and downloadable MIT-licensed weights.
The API economics are extraordinary.
| Model | Access | Context | Standard API price per 1M tokens, input / output | Practical position |
|---|---|---|---|---|
| DeepSeek V4-Flash | MIT weights + API | 1M | $0.14 / $0.28 | Exceptional high-volume value; smaller active model |
| DeepSeek V4-Pro | MIT weights + API | 1M | $0.435 / $0.87 | DeepSeek's strongest public model; especially competitive in code, maths, and long context |
| OpenAI GPT-5.6 Sol | Closed API | 1.05M | $5 / $30 | Top-tier general, coding, tool-use, and professional-agent performance |
| Anthropic Claude Fable 5 | Closed API | 1M | $10 / $50 | Anthropic's most capable broadly available model; strong in long-running coding and knowledge work |
| Google Gemini 3.1 Pro Preview | Closed preview API | 1.048M | $2 / $12 for prompts up to 200K | Strong multimodal reasoning, software engineering, and Google-grounded workflows |
Prices were checked on July 24, 2026. They are list prices, not the cost of completing an equivalent task; caching, reasoning effort, token use, latency, retries, tools, and provider discounts can change the result.
Even with that caveat, V4-Pro's uncached input price is about 11.5 times lower than GPT-5.6 Sol's, and its output price is about 34 times lower. V4-Flash extends the gap to roughly 36 times on input and 107 times on output.
That is DeepSeek's real attack.
It is not necessary for V4 to be the unambiguous best model if it is good enough for a large share of workloads at a small fraction of the price, with weights that customers can inspect and operate themselves.
The benchmark picture is more mixed.
DeepSeek's own V4 model card reports excellent coding and mathematics results and competitive scores on several agentic tests against the models available at its April launch. It also reports weaker results than the best closed competitors on some broad reasoning, terminal, tool-use, and professional-work evaluations. Those comparisons are already dated: GPT-5.6 and Claude Fable 5 arrived later, and the providers use different harnesses, reasoning budgets, and versions.
Current published frontier evaluations put GPT-5.6 Sol and Claude Fable 5 at the top of broad intelligence, coding-agent, and professional-work measures, with leadership changing by benchmark. Gemini 3.1 Pro remains a strong multimodal and tool-using frontier option. There is no credible single score that converts these systems into a permanent ranking.
The defensible conclusion is:
- DeepSeek is on the frontier in price-performance and open availability.
- It is highly competitive in coding, maths, and long-context work.
- The newest US models still have a meaningful edge on the hardest broad, long-horizon, and production-agent workloads.
- Deploying V4-Pro yourself is not “free” simply because the weights are open; a 1.6-trillion-parameter mixture-of-experts system demands serious infrastructure expertise.
Liang's own assessment is similarly restrained. He reportedly says DeepSeek still needs much larger active models to close the gap with the unpublished systems being developed in the US.
Open Source Is a Business Strategy, Not Charity
Liang argues that DeepSeek can publish its best weights and still make money because efficient deployment remains difficult. The company prices its API around a relatively short hardware-payback period and assumes that most third parties will not match its serving cost.
That logic is stronger than it first appears.
Open weights expand adoption, improve the surrounding ecosystem, create external optimization work, and make DeepSeek a default option for countries and companies that need control over deployment. The API then sells convenience and superior operations to customers who do not want to manage a giant model themselves.
This is the Red Hat insight applied to frontier AI: access to the artefact does not eliminate the value of operating it well.
The strategic contrast with the US is not simply open versus closed. Meta, open tooling, and a wide research ecosystem complicate that division. But DeepSeek is applying sustained pressure to the assumption that frontier capability must always carry frontier pricing.
What I Think Liang Gets Right
Liang is right that AI will reduce the human cost of hardware migration.
He is right that high-level kernel languages will make accelerator software more portable.
He is right that continuous learning—not another isolated benchmark record—is central to making agents genuinely useful inside organizations.
And he is right that cost, timing, and customer experience will matter more as model capabilities converge.
Where I disagree is the speed implied by the word “crumbling.”
CUDA lock-in is weakening. NVIDIA's full-stack advantage is not. The US is diversifying its compute base while simultaneously deploying NVIDIA's next platform. DeepSeek is proving that architectural efficiency and open weights can compress the economics of intelligence, while admitting that raw compute scale still limits what it can research and train.
The likely future is not one winner.
It is a bifurcated AI economy in which US labs lead the absolute capability frontier and system scale, while Chinese labs—DeepSeek in particular—push openness, portability, and price-performance much harder.
In that world, CUDA matters less as a language monopoly. NVIDIA may still matter enormously as the most complete way to turn electricity, chips, networks, and software into reliable intelligence.
The moat is not disappearing.
It is moving.
Sources
- The circulating English transcript includes the CUDA, AGI, open-source, pricing, and China-US comments and states that the text was not confirmed by Liang.
- DeepSeek's official V4 model card provides architecture, licensing, context, and vendor-reported evaluations; its API pricing page provides current prices and limits.
- The TileLang paper explains the higher-level tiled programming model behind Liang's portability argument.
- NVIDIA's official pages document TensorRT-LLM, Dynamo, the NVLink system layer, and the Rubin platform.
- Current model specifications and prices come from OpenAI, Anthropic, and Google. OpenAI's GPT-5.6 launch evaluations provide one current comparison among leading closed models.
- Alternative-accelerator evidence comes from official documentation for Google Ironwood, AWS Neuron and Trainium3, and AMD ROCm.