Backfire or Bite? US Chip Controls and China's AI, Explained

The answer, up front

Did US chip export controls backfire and accelerate China's AI? Not cleanly — but the counter-claim that they simply "worked" is also wrong. Both sides point at the same evidence — DeepSeek-V3 and R1, trained cheap, released open — and read opposite conclusions into it. This piece gives you the reading that survives scrutiny, plus a checklist to test any hot take you meet.

Here's the frame. DeepSeek's celebrated efficiency is a response to a constraint that was working, not proof the constraint was pointless. China's open-weight dominance is a deliberate distribution play, not a victory lap over Nvidia. And the whole "backfired vs. worked" fight is broken, because it treats a policy that changes every year as one fixed thing. The chips that trained DeepSeek-V3 were bought under a leaky 2022 regime, not the tighter 2024 one.

The person worth quoting here is Aravind Srinivas of Perplexity. On the 20VC episode "Micron Will Be More Valuable Than Meta | How Export Controls Helped Not Hurt China | Power is the Bottleneck to AI..." (published 2026-06-30, discussed at https://x.com/rohanpaul_ai/status/2071780740220752220), Srinivas makes several concrete claims that anchor this whole debate. He says the jury is still out on whether open Chinese models have caught up. He says that in the short term the controls are helping. He points to something like a 12-month gap between open-source releases and the closed frontier, and suggests the controls may be part of why that gap exists. And he says that because of the controls, DeepSeek is building around Huawei and local stacks — and that they made innovations in KV cache, attention, and training algorithms specifically to reduce memory and interconnect constraints. Even the frontier insiders aren't calling it settled. Neither should you.

Why a great cheap model doesn't mean the controls failed

The backfire story goes like this: we banned the good chips, China got creative, China shipped a frontier model for a few million dollars, therefore we handed them their AI industry. Each step sounds fine. The chain doesn't hold.

The move being smuggled in is conflating a demo with a scaled product, and an efficiency result with a strategic gift. DeepSeek being impressive tells you the team is good. It doesn't tell you the constraint was harmless. Those are different claims, and the whole debate lives in the gap between them.

Here's one thing to check tonight. Find any article claiming "sanctions backfired" and see whether it mentions interconnect bandwidth. If it only cites TFLOPS or the "$5.5M" figure, it isn't describing the actual bottleneck, and you can stop reading there.

What the controls actually choke

The policy doesn't ban "AI chips" as objects. It chokes three points where the US and its allies hold a near-monopoly. Understand these three and the whole debate reorganizes itself.

Leading-edge logic. Making the smallest transistors requires EUV lithography, and ASML — Dutch, allied — won't sell EUV machines to China. Without EUV you fall back to DUV multi-patterning: you expose the same wafer several times to draw features smaller than your light can normally resolve. Every extra exposure step multiplies the chances of a defect, so yield falls as die area grows. Be precise about where the wall is. 7nm is reachable on DUV — per teardown analyses widely reported by semiconductor research firms, SMIC's N+2 node fabs the Kirin phone chip that way, so "China can't do 7nm" is false. The wall is at high-volume, economical large dies, roughly 5nm and below. An AI accelerator die is enormous, and enormous area times a high per-step defect rate is what makes it uneconomical — not the node number itself.

Advanced packaging and memory. A modern accelerator isn't one chip. It's chiplets plus stacked HBM memory sitting on an interposer, assembled with CoWoS-class packaging. HBM (SK Hynix, Samsung, Micron) and packaging capacity are the real ceiling. You can have a working logic die and still ship nothing at volume, because you can't feed it or assemble it.

Interconnect on imported GPUs. The H800 was Nvidia's legal workaround: keep the raw math, cut the inter-GPU link bandwidth relative to the A100. The exact reduced figure for the H800 is reported inconsistently and depends on the variant, so treat it as "meaningfully reduced" rather than a hard number. This matters because large-model training is bottlenecked by moving data between GPUs — gradients, activations — not by the arithmetic itself. Throttle the link and you've reshaped what China can train, even with plenty of FLOPS on paper.

So the controls don't target the thing people argue about, raw compute. They target the three chokepoints — big-die yield, packaging and HBM, and interconnect — that decide whether you can scale and serve, not just demo. Hold that in your head. Every "backfire" hot take is really a claim about one of these three, usually without saying so.

Why the efficiency tricks prove the constraint bit

Now the part everyone gets backwards. Look at what DeepSeek actually did.

• Mixture-of-experts sparsity — activate only a slice of the model per token (roughly 37B active parameters out of 671B total in V3). Fewer parameters touched per step means less to synchronize across GPUs.
• FP8 mixed-precision training — use 8-bit numbers, halving the bytes you move and store versus 16-bit.
• DualPipe plus custom PTX kernels. DualPipe is a bidirectional pipeline-parallelism algorithm DeepSeek designed to overlap computation with communication — so while one stage is doing math, the interconnect is already moving the next batch's data, hiding the crippled link. Separately, they hand-wrote customized PTX instructions — below the CUDA abstraction — for the cross-node all-to-all communication kernels, squeezing the throttled bandwidth harder than the stock libraries would.

There's a broader pattern here, and it comes straight from the source. On the 20VC episode above, Srinivas says the DeepSeek team made innovations in KV cache, attention, and training algorithms explicitly to reduce memory and interconnect constraints. That lines up with what's in the public technical reports: cutting the size of the KV cache shrinks the memory footprint you have to hold and move per token, and reworking attention and the training procedure reduces how much you have to shuffle across a throttled link. This is not incidental. It's engineering aimed directly at the two things the controls squeeze — memory and interconnect.

Every one of these is a way to trade communication for computation. That's real, impressive engineering. It's also exactly what you'd build if someone handed you a chip deliberately spec'd to starve your interconnect. The optimizations are the fingerprint of the constraint.

Put the same team on unconstrained H100 clusters and they don't stop here — they go further, faster. "They did more with less" is not the same sentence as "the less didn't matter."

One more thing. That famous "$5.5M" cost is the marginal compute for one training run. It excludes the cluster capex, the prior research, and the failed runs. Quoting it as the total cost of building DeepSeek is like quoting the gas for one drive as the cost of owning a car.

Why Chinese labs give the weights away

"Sanctions pushed China to open-source, and open-source is winning" gets the causation wrong. Open-weighting isn't a cry of pain. It's a strategy with a name in tech: commoditize your complement.

A closed-API business needs two things: frontier chips at scale to serve inference, and open global sales channels. Export controls attack both for Chinese labs. So the rational move is to release the weights — the one asset that travels without a data center. Every fine-tune, tutorial, and downstream product built on Qwen or DeepSeek deepens dependence on the Chinese stack and quietly sets the tooling defaults. It's the Android playbook: give the OS away, own the ecosystem.

This connects to another point Srinivas makes on the episode: because of the controls, DeepSeek is building around Huawei and local stacks rather than assuming continued access to Nvidia. Open weights and a domestic-silicon roadmap are the same bet from two angles — reduce dependence on the thing you can't reliably import.

But notice that open-weight leadership and hardware sufficiency are two different axes. A model can top the leaderboard and have been trained on constrained, partly-rerouted Nvidia gear. Winning distribution doesn't mean the chips didn't bite. It means China played its remaining strong hand well.

Why "backfired" is a moving target (editorial analysis)

The following is LensUp.ai's own analysis, not a claim from the transcript. Here's the step almost every hot take skips. A rule's text and the physical flow of chips are not the same thing. There's always a lag between writing a rule and enforcing it — staffing, third-country routing, license loopholes. So you can't grade "the controls" as one object. You grade each cohort of chips against the enforcement regime that existed when those chips were bought.

The 2022 rule leaked. It opened a stockpiling window, and large volumes of H100/H800-class units are widely reported to have routed through Singapore, Malaysia, and reseller channels. On the Singapore signal specifically: Singapore's billed share of Nvidia's data-center revenue rose sharply against a small local install base, which is why it drew scrutiny — but Nvidia has stated in its filings that billing location is not the same as the delivery destination of the product. Read it as a flag worth pulling, not a proven smoking gun.

The 2023–2024 tightenings — an expanded Foreign Direct Product Rule, more Entity List additions, and closed de-minimis and license-exception gaps — materially raised the cost and friction of diversion. Not to zero. But a different regime.

The punchline: DeepSeek-V3 was trained largely on hardware acquired under the leaky regime. Judging the controls by that model measures 2022's enforcement gap, not whether the 2024 tightening works. People are grading the wrong exam.

The same logic applies to Ascend 910B. Its HBM was largely stockpiled or rerouted before the October 2023 rule. Separately, Reuters and US Commerce Department–linked investigations have reported probes into whether Ascend-class chiplets were fabbed via intermediaries before that supply was cut — reported as investigations, not as adjudicated findings. Treating those reports at their reported weight, the shape is still clear: cut those two inputs and you don't have a scaled product line, you have a launch event. Whether that's a two-year problem or a five-year moat is the honest debate.

The real open question

Strip away the noise and one genuine disagreement remains. Is the efficiency work — FP8, MoE, KV-cache and attention reductions, DualPipe and the PTX kernels — a durable capability China now owns and can carry onto domestic silicon? Or is it a one-time patch shaped exactly to Nvidia's crippled H800, that loses value the moment they retrain on Huawei's Ascend?

One view: the ceiling reappears at true frontier scale, because Ascend's yield and packaging limits are physical, not a matter of cleverness. The other: the transferable asset was never the chips — it's the human capital and tooling, and that doesn't un-learn itself.

The most defensible reading is that both are partly right, which is why "backfired: yes/no" is a category error. The controls bit — you can see the bite in the KV-cache and interconnect work Srinivas describes. They also trained a generation of Chinese engineers to build without abundant Nvidia bandwidth, which is a capability that outlives any single chip. Whether that second effect outweighs the first is a question about the next five years, not the last two. Notice that this is precisely the point the frontier insiders decline to call: Srinivas's "the jury is still out," paired with "short term the controls are helping," is not a dodge. It's an admission that the short-run and long-run signs point in opposite directions, and the crossover date is unknown.

Which side do you land on — durable capability, or a patch that loses its edge on Ascend? That's the fault line where the actual analysts disagree. Where you plant your flag should depend on one variable: how fast you think Ascend's yield and packaging constraints ease.

Do these three things before repeating either talking point

Go check the mechanism yourself instead of trusting the headline.

• Open any "sanctions backfired" piece and search it for "NVLink" or "interconnect." If the argument rests only on FLOPS or the $5.5M figure, it's skipping the bottleneck that actually shaped DeepSeek. Set it aside.
• Look up which enforcement year the hardware in question was acquired. Pre-October-2023 chips grade the leaky regime; don't let them stand in for the 2024 tightening.
• Separate the two axes on purpose. Ask whether the claim is about open-weight distribution or about hardware sufficiency — and refuse to let a win on one be sold as a win on the other.

Do those three, and the "backfire" question stops being a slogan and becomes a measurable one: which cohort of chips, under which regime, on which silicon roadmap. That's the version of the argument that can actually be settled — and it's the one worth having.