Nvidia got its H20 export license back on July 15, and AMD got the same clearance in the same breath, three months after the Commerce Department had killed it in April. That landed in the same fortnight the Senate killed a ten-year federal AI moratorium on a 99-1 vote, with Qualcomm’s Q3 earnings two weeks out and a data center bet attached that almost nobody took seriously a year ago. I am not going to pretend these three stories form a tidy thesis. They do not. But they landed on top of each other, and the order in which they matter is pretty clear.
The China reversal moves money this quarter. It goes first.
Export controls run through the Commerce Department’s Bureau of Industry and Security, the Entity List and the Export Control Classification Numbers on the Commerce Control List. In October 2022 BIS drew the first real line: anything above 600 TOPS aggregate transfer with a performance density past 10 TOPS per square millimeter needed a license. Nvidia engineered around it. The A800 and H800 throttled NVLink bandwidth from 600 GB/s down to 400 GB/s to slide under the bar.
October 2023 closed that door. The new test caught anything past 4,800 TOPS aggregate or 3.2 TOPS per square millimeter density, which killed the H800 outright. That is when the H20 was born, a Hopper die deliberately neutered to comply. The Biden-era AI Diffusion Rule in January 2025 then bucketed the planet into allied, restricted, and embargoed categories, placing China in the middle tier with a national compute cap measured in synthetic advanced computing units, TOPS times memory bandwidth over 1,000. BIS slapped a fresh licensing requirement on the H20 itself in April 2025, after intelligence assessments found that Chinese labs were lashing H20 clusters together over NVLink and training on them anyway, and the Trump administration scrapped the Diffusion Rule framework a month later, in May.
BIS has now issued specific validated licenses to both Nvidia and AMD to resume sales. The official BIS notice still has not been posted as of this writing, so treat the exact conditions as Bloomberg’s reporting rather than confirmed policy text.
The H20 is a weirder chip than its marketing suggests, and that is the thing people trip over. The compute got gutted, but the memory subsystem did not. It carries higher bandwidth than the full-fat H100 SXM5.
| Spec | H20 | H100 SXM5 |
|---|---|---|
| Architecture | Hopper (GH100, restricted) | Hopper (GH100) |
| FP16 Tensor | ~148 TFLOPS | 989 TFLOPS |
| INT8 Tensor | ~296 TOPS | 1,979 TOPS |
| Memory | 96 GB HBM3 | 80 GB HBM3 |
| Memory bandwidth | 4.0 TB/s | 3.35 TB/s |
| NVLink | 900 GB/s | 900 GB/s |
| TDP | 700W | 700W |
Read that bandwidth row twice. Nvidia shipped a China chip with roughly 15% of the H100’s tensor throughput but more memory bandwidth. For LLM inference, where you spend your life shuttling model weights out of HBM while the compute units sit half-idle, that profile is close to ideal. For training, which is compute-bound, it is hobbled. And yet bolt enough of them together over a fast fabric and the aggregate bandwidth starts to look like a training cluster. That is the exact loophole that got the H20 license killed in April.
AMD’s MI308X plays the same game from the CDNA3 side. Compute knocked down to about a quarter of the MI300X, memory left fully intact at 128 GB and 5.3 TB/s, the same regulatory logic and the same restored license as the H20.
The revenue stakes are not abstract. Nvidia pulled roughly $4.6 billion from China in FY2025, about 9% of total revenue, down from 21% two years prior, and Q1 FY2026 cratered after the April license requirement zeroed out H20 shipments. Analysts had modeled China data center revenue at basically nothing for the year. If these licenses cover H20 at volume, that is $4 to $6 billion swinging back onto the books, minus whatever volume caps BIS buried in fine print nobody has seen yet.
Huawei is the clearest loser. The Ascend 910C had been feasting on the export ban, and it is competitive with the H20 on raw FP16. Where it falls apart is memory: roughly 2.0 TB/s against the H20’s 4.0, which is the wrong place to be short when your customers are running inference. CANN and MindSpore are also a hard sell against a decade of CUDA codebases sitting in every Chinese lab. SMIC’s 7nm process, stuck on DUV lithography with no EUV access, yields nowhere near what TSMC’s 4N does, so Huawei cannot scale supply even when demand is there. The whole pitch was domestic supply with no export risk. Take away the export risk and the pitch collapses.
Where the Blackwell generation lands depends entirely on whatever thresholds BIS writes into the new framework. A China-market B20, if it follows the H20 playbook, would have its FP8 fused down to maybe 15-20% of the B200 while keeping the memory fat and the NVLink intact. Same trick, bigger die. Nvidia has gotten very good at clock-gating and firmware-locking exactly enough silicon to clear a customs threshold, and I would bet they have already taped out something for it.
The second story cuts the other way, and the headlines mangled it. The Senate voted 99 to 1 on July 1 to strip the federal AI moratorium out of the One Big Beautiful Bill Act, not to pass it. The House had inserted a ten-year preemption of state AI law back in May on a 215-214 knife-edge. The Senate spent late June fighting over its scope, a bipartisan amendment from Blackburn and Cantwell moved to kill it during the vote-a-rama, and it went down 99-1 with only Thom Tillis voting to keep it. The bill itself passed without the provision, the House took the Senate version on July 3, and the President signed it on July 4 with no moratorium in it. A 99-1 vote on anything is almost unheard of, and what it tells me is that a provision the entire tech industry lobbied hard for got radioactive the moment governors and state legislators on both sides actually read it.
Because it failed, the patchwork it would have frozen is still very much alive. The moratorium would have preempted the AI-specific laws: disclosure mandates, training-data rules, AI liability frameworks, algorithmic audit requirements, model-capability restrictions. Laws of general applicability that happen to touch AI were never in its blast radius, and they are not in question now either. California’s CCPA and CPRA privacy regime stands. Employment discrimination law applied to a hiring tool runs through the existing EEOC framework. Criminal law, professional licensing, state banking and insurance regulation, all untouched. The line everyone was fighting over was whether a statute names AI as the thing it regulates or just catches AI inside a broader net, and that line still matters, only for the opposite reason now: it is what every state AI law has to stay on the right side of to survive a future challenge.
Colorado’s SB 205 is the clearest example of what survives. Signed in May 2024, it is the most ambitious state AI law on the books, defining high-risk AI systems across employment, education, finance, healthcare, and housing, loading developers with risk and impact assessments, and handing the state AG civil-penalty authority. All of that machinery stays live and on schedule. California’s AB 2013, which forces anyone training on more than a million records to publish training-data documentation, is exactly the kind of AI-specific disclosure mandate the moratorium would have wiped, and it survives too. New York City’s bias-audit law for automated hiring tools was never really at risk, since the employment-discrimination angle put it outside the moratorium’s reach anyway.
For the engineering teams that build this stuff, the result is a headache that won’t go away, and I say that as someone with no appetite for handing the industry a ten-year regulatory holiday. A company deploying a national hiring model still maintains a separate compliance branch for each state that has passed a law. Colorado wants risk assessments and appeal mechanisms. NYC wants annual bias audits and candidate notification. Illinois has its video-interview consent rules. California layers AB 2013 on top of CCPA. Your audit logger still looks like a switch statement with a dozen state-specific branches, each writing to its own store, each with its own retention pipeline and disclosure UI. I have watched teams burn 200-plus engineering hours per model version on AB 2013’s training-data documentation alone: provenance records, consent and license tracking per data source, demographic representation analysis, the works. The industry lobbied to collapse all of that into one federal preemption, and on July 1 it lost. The switch statement stays.
The federal layer that was always there is unchanged. FCRA on credit and background decisions, FTC Act Section 5 on deceptive practices, HIPAA in healthcare, EEOC guidance in employment, plus the general-applicability state privacy laws. None of that was ever going anywhere. What did not happen is the AI-specific simplification the industry wanted, the one that would have cut compliance engineering for AI-specific work by something north of 60% overnight. That number was real in the pitch decks. It just is not happening.
The catch now is that this is not over. The amendment killed the moratorium in this bill, not the idea. The same coalition that lobbied for it is not going to walk away after one 99-1 loss, and the most likely path back is a narrower version attached to federal funding, a few years instead of ten, with carve-outs for the general-applicability laws that made this round politically toxic. So the planning posture is not relief, it is to build the multi-state compliance tooling properly now and assume the rules keep moving. Designing today on the bet that preemption is coming is how you end up rebuilding it twice.
The third story is the one I actually care about, and it is still unproven. Qualcomm reports Q3 FY2025 on July 30, fiscal quarter ended June 29, and the only line I am watching is whatever they are willing to say about data center. This is a company that lives on handsets and patent licensing trying to convince Wall Street it can sell inference silicon into the hyperscalers, and the pitch holds together better than the skeptics give it credit for. I am a declared Snapdragon Insider, so weigh my optimism with that in mind.
The whole bet rides on inference economics diverging from training economics. Training is compute-bound: FP16 and BF16, giant batches, GPU clusters running flat out. Inference is a different animal, memory-bandwidth-bound on the decode step, INT8 and INT4, small batches, and the metric that actually matters is dollars per query, which means power efficiency at scale. Qualcomm has spent a decade optimizing on-device inference for phones, where every milliwatt counts, and that institutional muscle is what the inference market rewards.
The CPU side is Oryon, the custom ARMv9.2-A core that Qualcomm acquired from Nuvia in January 2021 for $1.4 billion. It is a clean-sheet design, not a licensed Cortex, and the defining feature is cache. Oryon carries a 12 MB L2 that dwarfs the 2 MB on ARM’s own Neoverse V2, the core inside AWS Graviton4 and Google Axion. For small-batch LLM inference the payoff is direct: the KV cache for a single request can often live entirely in L2, so you never go out to DRAM and the bandwidth pressure that kills you on decode evaporates. The tradeoff is that Neoverse has years of hyperscale validation and a far more predictable profile across messy real-world workloads, which is not nothing when you are betting a data center on it.
Hexagon, the NPU that has ridden in Snapdragon silicon since the 820 in 2015, is the other half. The data center part is the AI 100 Ultra: four dies on a package, 128 GB of on-package LPDDR5X at roughly 4.0 TB/s, around 800 INT8 TOPS, a 150W TDP. Set that 150W against the H20’s 700W and the efficiency story writes itself, somewhere around 5.3 INT8 TOPS per watt against the H20’s 0.4. I will caveat that hard, because it is not a fair fight. The H20 carries training-adjacent baggage and the AI 100 Ultra is inference-only by design. But for the workload it is built for, the gap is real money.
I worked the inference pipeline on paper from the QNN SDK numbers, and the architecture makes sense. Unified LPDDR5X means no PCIe transfer tax, weights load once and stay accessible at full bandwidth, and the KV cache sits in the same pool. Quantize a 70B model to INT4 and it drops to roughly 35 GB, fitting comfortably in 128 GB with around 90 GB left for KV cache, enough to hold something like 180,000 concurrent tokens at INT8. On the estimated benchmarks, decode throughput lands near 800 tokens per second at batch 1, scaling to 12,000 at batch 32, at 150W. The H20 out-throughputs it at batch 1 thanks to that monster bandwidth, but it pays 700W to do it, so on tokens per second per watt Qualcomm comes out ahead by something like 1.8x.
Run the hyperscaler math and that efficiency turns into the only thing that matters at scale: cost per query. An H100 at thirty grand a unit burning 700W against an AI 100 Ultra at an estimated eight to twelve grand burning 150W changes the TCO equation, even though the Qualcomm part serves fewer queries per year. Memory bandwidth and watts decide inference deployments, not peak FLOPS, and that is the whole reason Qualcomm thinks it has a seat at this table. The company has already disclosed that an unnamed major hyperscaler committed to deploying AI 100 Ultra infrastructure, which is the validation the strategy needed. I will not speculate on who.
So on the 30th I am watching one thing above all: whether Qualcomm finally breaks out edge AI and data center as a real segment with real revenue attached, or whether it stays buried under IoT as a rounding error. Automotive will be strong again, almost certainly another record on the Snapdragon Digital Chassis ramp, and that part of the diversification story is more or less proven now. The data center line is the one still unproven, and the Snapdragon Summit, set for late September in Maui, is where the FY2026 narrative gets set, with next-gen Oryon cores and presumably an AI 100 Ultra successor on deck.
Pull back far enough and the three stories do rhyme. The H20 export license reopens China revenue, the domestic regulatory fight ends with the patchwork intact, and a credible non-Nvidia inference option starts shipping into the hyperscalers, all inside a few weeks. Where I land is uneven across the three. The China licenses read as pure trade leverage, and they will flip again the next time negotiations sour, so I would not build a product roadmap on them. The moratorium dying is the outcome I am most comfortable with, even with the compliance headache it preserves, because a ten-year regulatory blackout written by the people with the most to gain was never going to age well. Qualcomm is the one I am rooting for, because breaking Nvidia’s grip on inference is the rare outcome here that helps everyone who is not Nvidia. I have watched them overpromise on data center before, though, and shipping at hyperscaler volume is a different beast than shipping a press release.