China Is Now Copying Something Harder Than Nvidia's GPUs: CUDA

DeepSeek and Huawei have open-sourced compute, communication and programming tools for Ascend. The contest with Nvidia is moving beyond chip speed to how easily developers can switch.

Harder than building a GPU

AI developers do not work on a GPU's transistors directly. Their work runs through compilers, libraries, frameworks, kernels and communication software. Nvidia launched CUDA in 2006, and the edge it has built over roughly 20 years lies in that entire development environment.

That is why Nvidia's moat is better described as Hardware × Software × Developers. If a chip is fast but hard to program, and moving an existing PyTorch workload takes heavy engineering effort, total cost of ownership rises. The reverse also holds. If hardware is somewhat weaker but the software is easy to use and the whole cluster runs efficiently, a buyer's decision can change.

  • 2006: CUDA first released, followed by about 20 years of developer ecosystem building
  • 300+: CUDA-based libraries, according to Nvidia
  • 5M+: developers in the CUDA ecosystem, according to Nvidia

Three gaps DeepSeek filled

DeepGEMM-Ascend is a kernel library that runs the core matrix operations of AI efficiently on Huawei's Ascend chips. It supports BF16, FP8, FP4 and other formats, and keeps API compatibility with the original DeepGEMM. The point is that developers can keep their existing workflow while thinking less about Ascend's low-level hardware complexity.

DeepEP-Ascend is the more strategic piece. It handles communication among multiple NPUs, including expert-parallel traffic for mixture-of-experts (MoE) models. For very large AI models, performance and cost depend less on the FLOPS of a single chip than on how efficiently dozens, hundreds or thousands of accelerators are tied together.

TileLang's support for Ascend 950 lowers the barrier to entry. It is designed so developers can write high-performance kernels in a high-level, Python-like syntax while the tool handles Ascend code generation, scheduling and synchronization.

The three layers map to compute (DeepGEMM-Ascend, efficiency), communication (DeepEP-Ascend, chip-to-chip links) and programming (TileLang, ease of development).

China's answer is not only a better chip

Huawei's recent strategy makes the picture clearer. Built on UnifiedBus, the company links large numbers of NPUs in a single system. Over the longer term, it has laid out a roadmap to connect multiple SuperPoDs into clusters of up to 1 million NPUs. That figure is Huawei's plan. It does not mean the real performance and economics have been verified.

This is less a strategy of denying individual chips are behind than one of offsetting chip weakness with system scale and software.

That raises the value of communication software such as DeepEP-Ascend. The more accelerators are linked, the more latency, bandwidth and memory movement become bottlenecks. Owning thousands of chips does not help if effective compute time is low, and the economics break down. This is why the unit of competition in AI semiconductors is shifting from the chip to the effective throughput of whole racks and clusters.

Why DeepSeek, not Huawei alone

When a hardware vendor builds only its own tools, the result tends to be a closed, supplier-centered ecosystem. DeepSeek sits where real users do, training and running large models. That matters because the Ascend versions of DeepGEMM and DeepEP try to match existing APIs and development flows as closely as possible.

Here, open source works as a distribution strategy. Outside developers use the tools, find bugs and add optimizations, which draws in more developers: a developer flywheel. What China is trying to copy is less the CUDA code itself than the network effect CUDA has built over the past 20 years.

What investors should watch

The key question is whether these tools draw real developer adoption, and whether Ascend clusters deliver competitive effective throughput and economics. Huawei's 1-million-NPU roadmap is a plan, not a verified result.

China is likely aiming less at a faster chip than at an environment where developers can leave Nvidia more easily.

Insight Times Editorial Desk

More by topic

Starting with this story's topics, then what readers are following now

Investor psychology, decision-making and books on investing

NVDA

This story's topic101 storiesSee all →

Why Nebius Is Keeping Some 2027 AI Capacity Off the Shelf

Nebius mixes long-term contracts, short-term premiums and auctions. Holding back capacity pays only if the servers stay busy, and it is not yet clear the pricing will last.

Tech

Who Hits $1 Trillion Next: Reading AMD and ASML by Different Rules

One company is chasing Nvidia's GPU empire. The other sells the machine that draws the chips for every advanced chipmaker, including Nvidia. Both get floated as 2028 trillion-dollar candidates, but the numbers investors need to watch are entirely different.

Investing

AI ran out of memory before it ran out of GPUs

The HBM bottleneck is spreading into DRAM prices, server costs and the price of the next laptop or phone.

Tech

Box Shows AI Its Manuals, and the AI Starts Redesigning the Workflow

Box Architect, built with NVIDIA, moves AI from reading documents to designing how work gets done. It has not yet reached beta, so the real-world payoff remains to be seen.

Tech

Tech Selloff: Is the Bull Market Over? Wall Street's Four-Year Chart Says Not Yet

The uptrend has survived five double-digit pullbacks in four years. This time, AI companies' debt and cash flow are the new variables.

Investing

Musk Sees the Next Phase of AI Infrastructure: Who Can Switch GPUs On, Fast

If neoclouds struggle under financial pressure, AI demand may not shrink. It may simply favor the few operators with the capital and execution to serve it, a shift that could matter for SpaceX and Tesla.

Tech

Marvell's $115 Billion Lumentum Deal: Is AI's Next Bottleneck Light, Not Chips?

Marvell's purchase of Lumentum points to a contest beyond GPU speed, in data transfer and power efficiency. Whether the price is right is not yet confirmed.

Tech

GPU

This story's topic73 storiesSee all →

A chip built to draw graphics, whose ability to run thousands of identical cal…

Is Space Cooler for AI? Why Musk Wants Data Centers in Orbit

The next limit on AI may be power and cooling, not GPUs. Elon Musk is looking to Earth orbit for an answer, but the test is cost per hour of compute.

Tech

Nvidia and Palantir Put AI Inside the Supply Chain - The Real Shift Is That It's Learning Judgment

A single Vera Rubin rack holds 1.3 million parts. The real breakthrough in Nvidia and Palantir's new supply chain AI isn't automating that complexity, it's capturing the judgment calls that used to live only in planners' heads.

Tech

The Next AI Chip Race Is About How Chips Connect

Applied Materials will lift its stake in Dutch equipment maker BESI from about 9% to 20%. The bet is less about control than about a new way to link chips.

Tech

Nvidia Moves From GPUs to CPUs as AI Agents Reshape a $200 Billion Chip Market

A customer's early test found Nvidia's Vera CPU handled about twice the AI agent workload of its existing chips. The bigger story is where the money in semiconductors is heading as AI shifts from answering questions to doing work.

Tech

US AI Data Centers Are Headed for 90GW. Gigawatts Now Matter More Than Dollars

Goldman Sachs expects US data center capacity to grow from 64GW at the end of 2026 to 90GW at the end of 2027. The AI race is moving from buying GPUs to building power plants and transmission lines alongside them.

Tech

Nvidia's $500 Billion Experiment: Wall Street Asks Who Backs the GPU, Not Just the Chip

The real issue in Nvidia's financing platform is not chip sales. It is how far Wall Street will trust GPU resale prices and rental income three to four years out.

Tech

The Cheaper Tokens Get, the Bigger Your AI Bill Gets

As AI shifts from training to inference, the token is becoming an industrial commodity. Falling prices are not cutting demand, because agents and automation consume tokens at a scale no human chat session ever did.

Investing

Semiconductors

This story's topic61 storiesSee all →

The whole business of designing, fabricating and testing chips.

A 3x P/E is not a discount. It is a verdict on how long memory profits last

Nomura's buy case on Samsung Electronics and SK Hynix does not stop at higher memory prices. It argues that if AI has turned memory from a cyclical component into strategic infrastructure, the formula the market uses to value both companies has to change too.

Investing

The Real ASML Number Isn't the 3% Pop, It's the 30% Capacity Hike

ASML's rally isn't just an AI trade. Customers are already locking in 2027 tool slots, and ASML is raising Low-NA EUV capacity by roughly 30% to keep up.

Investing

White House Swaps AI for "SI." The Real Investment Signal Is Infrastructure, Not the Name

An executive order renames artificial intelligence "Super Intelligence" and six frontier labs sign a voluntary safety accord. What investors should watch is the data center, power, nuclear and chip policy now gathered under one national-competitiveness frame.

Tech

Samsung Posts 107 Trillion Won Profit, TSMC Sales Jump 50%. Why Did the Shares Fall?

The AI chip supercycle has not ended. But the market's question has changed: not how fast earnings grow, but how long today's enormous margins can last.

Tech

China's AI Labs Went Open Source Because They Had to, Not Because They Wanted To

A GPU shortfall pushed Chinese AI labs toward efficiency and open weights, and that constraint is now reshaping the global competition over who controls AI's basic infrastructure.

Tech

Stocks Bounce, but the AI Selloff May Not Be Over: A Two-Week Playbook

The Dow rose more than 1% while the Nasdaq barely moved and semiconductors kept falling. A W-shaped retest looks more likely than a V, and the task now is surviving the next drop.

Investing

Memory Chip Supercycle: Not at Peak Yet, But 5 of 6 Warning Signs Are Flashing

Memory prices are still climbing, but the pace is slowing and PC and smartphone demand is starting to crack under the weight of higher chip costs. This looks less like a price peak and more like a profit plateau building toward the next supply cycle.

Investing

Chinese Economy

This story's topic22 storiesSee all →

Weekly Check: The AI Bull Market Isn't Over, It's Just Gotten Harder

A 5% 10-year yield, 470 gigawatts stacked up in Texas, Chinese DRAM, and whether data centers actually get switched on are starting to decide who wins the AI trade through 2027.

Tech

AI's Next Bottleneck Is Power, Not GPUs

Jensen Huang and Elon Musk sent the same warning in Washington. The AI race is widening from building better models to supplying power and chips at industrial scale.

Tech

What Matters More Than Tesla Semi's 2,500-Truck Order

Tesla was picked as lead supplier in the largest electric Class 8 truck group order in US history. The bigger story is that the fight in electric trucking is shifting from range to charging, financing, uptime and autonomy.

Tech

The US-China Summit Didn't Erase Taiwan Risk for Chip Stocks

Washington's Taiwan policy held, but roughly $14 billion in arms sales is still on hold. For AI chip investors, the summit reads less like risk removed and more like a fragile balance extended.

Macro

Lam Research Falls 12%: The Problem Is the Timing of AI Growth, Not Its End

Revenue up 54%, an operating margin above 40% and $3.2 billion in free cash flow were not enough. The selloff likely reflects worry that revenue arrives later than hoped.

Investing

China Wants to Replace Nvidia, but Its Tech Giants Ordered $35 Billion of Its Chips

A huge order for the RTX PRO 5500 shows where AI inference demand stands, and what Nvidia can and cannot turn into revenue.

Tech

AI Won't Slow Down. The Real Fight Is Over What Agents Can Do

Dario Amodei wants frontier AI labs to slow capability gains, but a US-wide deceleration deal looks unlikely. The more probable path: keep accelerating commercial AI while tightening security around agent permissions, high-risk features, model weights and large-scale compute access.

Tech