# The US–China AI Chip War and Export Controls — Reading BIS's EAR, ECCN 3A090 and TPP from the Rule Text and Company Filings (as of November 17, 2026)

> US export controls on AI chips to China, read from the rule text and NVIDIA's SEC filings: ECCN 3A090, TPP, the 2022–2026 rules, H20, H200 and Japan's controls.

- Published: 2026-11-17
- Author: 友田 陽大
- Tags: AI, セキュリティ, インフラ, 技術選定
- URL: https://tomodahinata.com/en/blog/us-china-ai-chip-export-controls-bis-ear-explained
- Category: NVIDIA, GPUs & AI chips
- Pillar guide: https://tomodahinata.com/en/blog/nvidia-gpu-ai-tensor-core-fp8-hbm-nvlink-explained

## Key points

- US controls are not a blanket ban on China. They require a license to export AI chips that fall under ECCN 3A090 and related entries to China, Macau and other destinations, and they set different review policies such as "presumption of denial" and "case-by-case".
- The line is drawn by TPP (total processing performance = dense peak throughput × bit length) and performance density (TPP ÷ die area). By the author's calculation, the H100 SXM has a TPP of 15,831 and the H200 15,832 — both under the TPP 21,000 ceiling the January 2026 rule set for case-by-case review.
- The 2022 rule also required a transfer rate of 600 GB/s or more, so the A800 and H800, with NVLink capped at 400 GB/s, could fall outside it. The 2023 amendment removed that condition from the item definition, and TPP alone now catches them.
- NVIDIA's filings say the H20 license requirement led to a USD 4.5 billion charge in Q1 of fiscal 2026, and that even with H200 licenses, Chinese government restrictions kept sales down and left it "effectively foreclosed" from China's data center market.
- The January 2025 AI Diffusion Rule was never enforced but is still in the regulations. The rule that formally rescinds it is scheduled to appear in the Federal Register on November 19, 2026 (not yet in force on November 17). This article is not legal advice.

---

News coverage tends to compress US export controls on AI chips into one phrase: "a ban on exports to China." The rule text is more detailed than that, and it has changed many times.

Here is the short answer.

- The US controls are **not a blanket ban**. They **require a license** to export chips above certain performance levels to China, Macau and other destinations. Whether a license is granted depends on a **review policy** that varies by item and destination.
- The line is drawn by two numbers: **TPP (total processing performance)** and **performance density**. The regulation gives a formula for both, so you can estimate them from published specs.
- The rules were amended repeatedly between October 2022 and November 2026. Watch for **publication dates that differ from effective dates**, and for **restrictions imposed by individual notices rather than published rules** (the H20 is one).
- As of November 17, 2026, the January 2025 "AI Diffusion Rule" remains in the regulations, unenforced. The rule that formally rescinds it is **scheduled to be published in the Federal Register on November 19** and is not yet in force.

**About this article.** It is a technical explainer based on primary sources I could check as of November 17, 2026 (Japan time). It is **not legal advice**. For any real export, reexport or transaction, consult an export control specialist or a lawyer. I keep opinions about whether the policy is good separate from facts about the rule text and numbers, and I name the source of any opinion I report.

| Label in this article | Meaning |
|---|---|
| Rule text | The Federal Register and the Code of Federal Regulations (CFR), quoted verbatim |
| Company filings | Documents NVIDIA and AMD filed with the US Securities and Exchange Commission (SEC), and official product specs |
| Reporting | News articles, attributed as "X reported" |
| Author's calculation | Values I computed from published specs using the regulation's formula. Not an official classification |

## 1. The short answer: what is controlled and what is not

Three common misconceptions come first.

**Misconception 1: "NVIDIA cannot sell GPUs to China."** NVIDIA's quarterly report (10-Q) for the second quarter of fiscal 2027 (the quarter ended July 26, 2026) says this about uncontrolled products:

> while we were able to ship uncontrolled products to China, such as gaming and workstation GPUs, we were effectively foreclosed from competing in China's data center computing/compute market

So the company says it can still ship gaming and workstation GPUs, but has been "effectively foreclosed" from the data center market. It attributes that not only to US rules but also to the Chinese government discouraging purchases (Section 7).

**Misconception 2: "A license requirement means a ban."** A license requirement means you may not export without authorization. Whether authorization is granted depends on the **review policy**. The EAR uses three main levels:

| Review policy | Meaning |
|---|---|
| Presumption of denial | Generally denied, with some room for exceptions |
| Case-by-case | Decided transaction by transaction |
| Presumption of approval | Generally approved |

**Misconception 3: "The controls are a single law."** In practice, many interim final rules (IFRs) from BIS are stacked on top of each other, combined with notices to individual companies, guidance documents and a presidential tariff proclamation. If you do not track which document and which date a claim refers to, you will mistake old information for current information.

## 2. Why AI chips became a national security issue: a supply chain map

To understand what the controls aim at, you need to know how AI chips are made. Think of the controls not as one gate but as several: a **finished-product gate**, a **factory gate** and a **parts gate**.

| Stage | Main players (as far as I could confirm) | Relevant US controls |
|---|---|---|
| Design | NVIDIA, AMD; in China, Huawei (Ascend) | Classification of finished products (ECCN 3A090, 4A090, .z items) |
| Front-end manufacturing (foundry) | NVIDIA's 10-K says it uses TSMC and Samsung | Due diligence duties for foundries (January 2025 rule) |
| Manufacturing equipment | ASML (Netherlands) for lithography; Japanese equipment makers cover many process steps | Equipment classification (ECCN 3B001 and others), foreign direct product rules |
| Memory (HBM) | NVIDIA's 10-K says it buys from SK hynix, Micron and Samsung | HBM classification (ECCN 3A090.c, December 2024 rule) |
| Advanced packaging | NVIDIA's 10-K says it uses CoWoS | — |

NVIDIA's annual report (10-K) for fiscal 2026 describes its supply chain this way:

> We utilize foundries, such as Taiwan Semiconductor Manufacturing Company Limited, or TSMC, and Samsung Electronics Co., Ltd., or Samsung, to produce our semiconductor wafers. We purchase memory from SK Hynix Inc., Micron Technology, Inc., and Samsung. We utilize CoWoS technology for semiconductor packaging.

Why control so many stages? Blocking only finished chips does little if the other country can obtain the equipment and build its own. Blocking only equipment does little if finished chips can still be bought. The US rules cast a net over finished chips, contract manufacturing, equipment and memory.

The lithography rule text shows where the technical choke point is. ECCN 3B001.f.1 covers lithography tools with a light source shorter than 193 nm (EUV, at 13.5 nm, qualifies) and tools at 193 nm or longer that reach a certain resolution and overlay accuracy. 3B001.f.1 is on the list of items that need a license for China, Macau and the other listed destinations (§742.6(a)(6)(i)(A)).

Why GPUs are so strong at AI (Tensor Cores, low-precision math, HBM, NVLink) is covered in detail in [Why NVIDIA GPUs are so strong at AI](/blog/nvidia-gpu-ai-tensor-core-fp8-hbm-nvlink-explained). Terms used here such as "dense vs sparse," "FP8" and "HBM bandwidth" assume that background.

## 3. EAR basics: ECCNs, license requirements, review policies and license exceptions

Before reading the rule text, let us agree on the minimum vocabulary.

- **EAR (Export Administration Regulations)**: the US regulations that control exports of commercial and dual-use items, at 15 CFR 730–774.
- **BIS (Bureau of Industry and Security)**: the Commerce Department bureau that administers the EAR.
- **CCL (Commerce Control List)** and **ECCN (Export Control Classification Number)**: the list of controlled items and their classification numbers. AI chips are mainly classified under "3A090".
- **Country groups**: the EAR sorts countries into groups. D:1, D:4 and D:5 are groups of national security concern, and China is in D:5.
- **Headquarters rule**: since the 2023 amendment, a license is required when the recipient is **headquartered in, or has an ultimate parent headquartered in,** a D:5 country or Macau, even if the recipient itself sits in another country.
- **License exceptions**: conditions under which you can export without an individual license. For AI chips these include NAC (Notified Advanced Computing), which requires prior notice to BIS, and License Exception HBM for high-bandwidth memory.
- **End-use controls**: controls keyed to how an item will be used rather than what it is. §744.23, which prohibits exports with knowledge that they are for supercomputer and similar end uses, is the main example.
- **Foreign direct product (FDP) rules**: provisions that extend the EAR to foreign-made items produced from US technology or software.
- **General Prohibition 10 (GP10)**: prohibits dealing with an item when you know a violation has occurred or is about to occur.

## 4. Reading the text of ECCN 3A090, 4A090 and the .z items

### 4.1 The text of 3A090

The current text (eCFR, updated as of November 12, 2026) defines 3A090.a as follows:

> a. Integrated circuits having one or more digital processing units having either of the following: a.1. A 'total processing performance' of 4800 or more; or a.2. A 'total processing performance' of 1600 or more and a 'performance density' of 5.92 or more.

3A090.b is the tier below:

> b.1. A 'total processing performance' of 2400 or more and less than 4800 and a 'performance density' of 1.6 or more and less than 5.92, or b.2. A 'total processing performance' of 1600 or more and a 'performance density' of 3.2 or more and less than 5.92.

In table form:

| Tier | Condition |
|---|---|
| 3A090.a | TPP ≥ 4,800, or TPP ≥ 1,600 and performance density ≥ 5.92 |
| 3A090.b | 2,400 ≤ TPP < 4,800 and 1.6 ≤ performance density < 5.92, or TPP ≥ 1,600 and 3.2 ≤ performance density < 5.92 |
| Excluded (Note 2) | Not designed or marketed for use in datacenters, and TPP below 4,800 |

Read Note 2 carefully. A product that is not for datacenters **still falls under 3A090 if its TPP is 4,800 or more**. In that case the text points to License Exceptions NAC or ACA. This is why gaming GPUs come up in export control discussions.

### 4.2 Definitions of TPP and performance density

The definition of TPP (Technical Note 2 to 3A090) reads:

> 'Total processing performance' ('TPP') is 2 × 'MacTOPS' × 'bit length of the operation', aggregated over all processing units on the integrated circuit.

As a formula:

$$
\mathrm{TPP} = 2 \times \mathrm{MacTOPS} \times \text{bit length}
$$

MacTOPS is the theoretical peak of multiply-accumulate operations (D = A × B + C), in trillions per second. Industry convention counts one multiply-accumulate as two operations, so the text itself says that "2 × MacTOPS may correspond to the reported TOPS or FLOPS on a datasheet." In practice you can compute:

$$
\mathrm{TPP} \approx (\text{dense peak TOPS or TFLOPS on the datasheet}) \times \text{bit length}
$$

The technical notes add several rules:

- The bit length is "the largest bit-length of the inputs to the multiply operation": 8 for FP8, 16 for FP16.
- Use the highest value the manufacturer claims. If the chip supports several bit lengths, **take the highest TPP**.
- Even if the chip supports sparsity, **use the dense value** ("the values for processing of dense matrices (e.g., without sparsity)").

The text's own example: 600 trillion operations at 8 bits, or 300 TFLOPS at 16 bits, gives exactly 4,800.

Performance density is TPP divided by die area:

$$
\mathrm{PD} = \frac{\mathrm{TPP}}{\text{applicable die area [mm}^2]}
$$

"Applicable die area" is the total area of all logic dies made on a process with non-planar transistors (such as FinFET).

Why is performance density needed? The 2023 rule explains:

> A performance density parameter prevents the workaround of simply purchasing a larger number of smaller datacenter AI chips which, if combined, would be equally powerful as restricted chips.

Buy enough small chips and gang them together, and you match a controlled large chip. Adding "performance per unit area" closes that loophole.

### 4.3 3A090.c (HBM), 4A090 and the .z items

The December 2024 rule added 3A090.c, which covers HBM (high-bandwidth memory). The condition is a "memory bandwidth density" greater than 2 GB/s per square millimeter, where memory bandwidth density is bandwidth (GB/s) divided by the area of the package or stack (mm²). A co-packaged device that combines logic with HBM and whose dominant function is processing (that is, the GPU itself) is outside 3A090.c.

4A090 covers computers and similar items that incorporate a 3A090 chip. The ".z items" (3A001.z, 4A003.z, 4A004.z, 4A005.z, 5A002.z, 5A004.z, 5A992.z, 5D002.z, 5D992.z) pick up products that incorporate a chip meeting the 3A090 parameters within their own classifications. Think of them as the mechanism that keeps a GPU server or network appliance from slipping through under its original classification.

## 5. Timeline: publication, effective dates, scope and review policy

Here are the main rules I confirmed in the Federal Register. "Published" is the date of publication in the Federal Register; "effective" is the date the rule takes effect. Some rules took effect before they were published.

| Document | Published | Effective, etc. | Scope (key points) | Review policy / exceptions (key points) |
|---|---|---|---|---|
| Oct 2022 IFR (87 FR 62186) | 2022-10-13 | Parts on 2022-10-07; 3A090 and 4A090 added on 2022-10-21 | Created 3A090 and 4A090. License requirement for China. Supercomputer and advanced-IC end-use controls, FDP rules, controls on US persons' support | Most applications: presumption of denial. Equipment for China fabs of companies headquartered in the US and similar countries: case-by-case |
| Oct 2023 advanced computing IFR (88 FR 73458) | 2023-10-25 | 2023-11-17 | Reorganized 3A090 into .a/.b (TPP and performance density). Created .z items. Expanded to D:1, D:4, D:5 (excluding A:5, A:6). Added D:5/Macau-headquartered entities | Presumption of approval outside D:5/Macau (except for D:5/Macau-headquartered entities); presumption of denial for D:5/Macau. New License Exception NAC |
| Oct 2023 equipment IFR (88 FR 73424) | 2023-10-25 | 2023-11-17 | Expanded equipment controls | — |
| Apr 2024 corrections (89 FR 23876) | 2024-04-04 | 2024-04-04 | Corrections and clarifications to the 2023 rules | — |
| Dec 2024 FDP/HBM IFR (89 FR 96790) | 2024-12-05 | 2024-12-02 (compliance for HBM and FDP: 2024-12-31) | New equipment controls, new FDP rules, new HBM entry (3A090.c) | HBM needs a license to D:5/Macau. New License Exception HBM |
| Jan 2025 AI Diffusion Rule (90 FR 4544) | 2025-01-15 | 2025-01-13 (compliance 2025-05-15; some items 2026-01-15) | Worldwide license requirement for 3A090.a and others. Country caps, controls on model weights (4E091) | New License Exceptions AIA, ACM, LPP. 18 countries including Japan, plus the US, got favorable treatment |
| Jan 2025 foundry due diligence IFR (90 FR 5298) | 2025-01-16 | 2025-01-16 (compliance 2025-01-31) | Requires foundries and back-end firms to check advanced logic ICs before shipment | Approved IC designer mechanism |
| May 2025 rescission announcement, GP10 guidance | 2025-05-12 to 13 | — | Announced the start of rescission and non-enforcement. Warned about the risks of using Huawei Ascend | — |
| Sep 2025 VEU revocation (90 FR 42321) | 2025-09-02 | 2025-12-31 | Removed the China sites of Intel (Dalian), Samsung and SK hynix from the Validated End-User list | — |
| Sep 2025 Affiliates Rule (90 FR 47201) | 2025-09-30 | 2025-09-29 | Companies at least 50% owned by Entity List parties automatically get the same restrictions | Stayed from 2025-11-10 (90 FR 50857); stay extended to 2027-01-10 (91 FR 70789) |
| Jan 2026 review policy revision (91 FR 1684) | 2026-01-15 | 2026-01-15 | Exports from the US to China/Macau of products with TPP < 21,000 and DRAM bandwidth < 6,500 GB/s (e.g., H200, MI325X) | From presumption of denial to case-by-case if conditions are met |
| Jan 2026 Proclamation 11002 (91 FR 2443) | 2026-01-20 | Tariff from 2026-01-15 | 25% tariff on imports of certain advanced computing chips (US data centers and other uses excluded) | — |
| Jul 2026 UAE rule (91 FR 43034) | 2026-07-14 | 2026-07-10 | License-free supply of advanced items to the UAE government, approved UAE entities and US-headquartered AI companies | — |
| Nov 2026 remote access IFR (FR Doc 2026-23529) | **Scheduled 2026-11-19** | Effective on publication | Controls compute provided through data centers. Adds five countries. Formally rescinds the AI Diffusion Rule | New licenses reviewed case-by-case |

Some details follow.

### 5.1 October 2022: the first framework

The rule published on October 13, 2022 created 3A090 and 4A090. Effective dates varied by provision: adding 3A090 and 4A090 to the CCL took effect on October 21, and part of the end-use controls on October 7. Note that some parts took effect before publication.

The 3A090.a of that time was structured differently:

> a. Integrated circuits that have or are programmable to have an aggregate bidirectional transfer rate over all inputs and outputs of 600 Gbyte/s or more to or from integrated circuits other than volatile memories, and any of the following: a.1. … bit length per operation multiplied by processing performance measured in TOPS, aggregated over all processor units, of 4800 or more;

In other words, a chip was covered only if it met **both** conditions: an **aggregate bidirectional I/O transfer rate of 600 GB/s or more**, **and** "bit length × TOPS" of 4,800 or more. That "and" is what leads to the A800 and H800 story (Section 6).

### 5.2 October 2023: performance density and wider country scope

The rule published on October 25, 2023 and effective November 17 deleted 3A090.a and replaced it with the new .a and .b. The new text has no transfer-rate condition. Classification now rests on compute (TPP) and performance density alone.

The country scope widened as well, from "China and Macau" to D:1, D:4 and D:5 countries (excluding those also in A:5 or A:6), and to entities headquartered in D:5 or Macau wherever they are. Review policy: presumption of denial for D:5/Macau, presumption of approval elsewhere (except for D:5/Macau-headquartered entities).

### 5.3 December 2024: HBM and equipment

The rule published on December 5, 2024 took effect on December 2. But the compliance date for HBM (3A090.c), the new FDP rules and some other parts was December 31. This is an example where the "effective date" and the "compliance date" differ.

### 5.4 January 2025: the AI Diffusion Rule and what came after

The "AI Diffusion Rule," published January 15, 2025, took effect on January 13 with a compliance date of May 15. It imposed a worldwide license requirement on 3A090.a and related items, capped compute per country (790,000,000 TPP over 2025–2027), and even controlled AI model weights. Japan was on the list of 18 favored countries (plus the US).

In a press release dated May 12, 2025, BIS said it had "initiated" a rescission and instructed enforcement staff not to enforce the rule. The release said BIS would publish a Federal Register notice formalizing the rescission and issue a replacement rule in the future.

The key point: **as of November 17, 2026, the AI Diffusion Rule's text is still in the CFR**. The eCFR entry for 3A090 still says "RS applies to 3A090.a: To or within any destination worldwide," and still lists License Exceptions AIA and LPP. The text is unenforced, not deleted.

The formal removal comes in FR Doc 2026-23529, "Addressing Remote Access to Advanced Semiconductors." This interim final rule went on Federal Register public inspection on November 16 and is **scheduled to be published and take effect on November 19**. According to the public inspection copy, it does four main things:

1. A new end-use control (§744.24). It requires a license to export advanced items with knowledge that they are for a data center (or a set of linked data centers) that will give companies headquartered in D:5 or Macau access to more than 12,000,000 "aggregate TPP." Use solely for commercial recommender systems is excluded, provided the reporting requirements are met. For data centers already providing such services, the preamble says in effect that they must obtain a license within 90 days of the effective date or face future export restrictions. How they are treated during those 90 days is not clear from the public inspection text.
2. A license requirement for advanced items to Indonesia, Malaysia, Singapore, Brazil and Thailand. Consignees headquartered in the US are excluded if reporting conditions are met.
3. Fewer ways for foundries and back-end firms to show that a chip they ship is outside the controls.
4. Formal rescission of the AI Diffusion Rule. It removes the model-weights control (4E091), the country tiers, License Exceptions AIA and LPP, and the Data Center VEU, while keeping License Exception ACM.

A footnote in the rule says that, under Executive Order 14434 (signed September 29, 2026; published October 2), it uses "Super Intelligence (SI)" in place of "Artificial Intelligence (AI)." The wording in the regulation changed; the technology it covers did not.

### 5.5 The H20: a notice, not a rule

The H20 restriction is not a Federal Register rule. According to NVIDIA's 10-K, it is a license requirement the US government **notified NVIDIA of** in April 2025:

> In April 2025, the USG informed us that it requires a license for export to China (including Hong Kong and Macau) and D:5 countries, or to companies headquartered or with an ultimate parent therein, of our H20 integrated circuits and any other circuits achieving the H20's memory bandwidth, interconnect bandwidth, or combination thereof.

Note that the criteria are **memory bandwidth and interconnect bandwidth**, not 3A090's TPP or performance density. A specific product can be restricted by criteria that do not appear in the published regulations. AMD disclosed a new license requirement for its MI308 in an 8-K at the same time (Section 7).

### 5.6 January 2026: H200-class chips move to case-by-case

The rule published and effective on January 15, 2026 changed the review policy for exports from the US to end users in China and Macau, from presumption of denial to case-by-case, for products meeting these conditions:

> commodities with a TPP less than 21,000 … and a 'total DRAM bandwidth' less than 6,500 GB/s … such as the NVIDIA H200 or AMD MI325X

Only products commercially available in the US when the rule was published qualify. To get case-by-case review, the applicant must also certify, among other things:

- Supply to US customers is sufficient and US orders will not be delayed; foundry capacity will not be diverted to China-bound production
- The aggregate TPP shipped to China and Macau is **no more than 50%** of the aggregate TPP of the same product shipped to US customers
- The transaction is not for prohibited military or similar end uses or users; the consignee runs know-your-customer (KYC) checks, and remote-access users are disclosed
- Every shipment is checked by an **independent third-party lab headquartered in the US, with testing done in the US**

Reexports to D:5/Macau, and exports to entities headquartered in D:5/Macau wherever they are, remain under a presumption of denial. A headline like "H200 unbanned" drops all of these conditions.

## 6. Original analysis: computing TPP from published specs

Now I apply the regulation's formula to compute TPP for some major chips. **This is not an official classification.** It does not replace the manufacturer's classification, a BIS determination or expert review.

### 6.1 Assumptions

- Figures are **dense** peaks from each vendor's official datasheet or product page. Where only a sparsity figure is given, I halved it (the H100 whitepaper and Blackwell datasheet state that dense is half of sparse).
- The H800 figures come not from NVIDIA but from a product guide by the server maker Lenovo. The A800 figures are for the "A800 40GB Active" workstation card that NVIDIA documents publicly, which is a different product from the datacenter A800 sold in China. Its product page lists "1,247 AI TOPS" (INT8, sparse) and "623.8 TFLOPS"; the latter has no bit length given, so I assumed it is the FP16 sparse figure.
- The L40's 608.5 mm² die area is my inference: the whitepaper's L40 table gives "GPU Code Name: AD102," and the AD102 table gives that area. The L40 table itself gives no area.
- The 2022 column approximates the old definition (bit length × TOPS, and a transfer rate of 600 GB/s or more) using the same datasheet values. The H200 and B200 did not exist in 2022, so that column is only a "what if" reference.
- For the transfer rate I simply added published NVLink and PCIe figures. How to count "all inputs and outputs" is a classification judgment, so this too is my simplification.
- I did not compute the H20 because I could not find official NVIDIA specs for it.

### 6.2 Code

Environment: Python 3.9 or later (standard library only). I ran it with Python 3.9.6 on an Apple M4 Pro (macOS). Run it with `python3 tpp_check.py`.

```python
# tpp_check.py — author's estimate. Not an official classification.
# Formula source: 15 CFR part 774 supplement no. 1, ECCN 3A090, Technical Note 2
#   TPP = 2 × MacTOPS × bit length = datasheet dense TOPS/TFLOPS × bit length
#   Compute for each bit length and take the maximum. Never use sparsity figures.
#   Performance density (PD) = TPP ÷ applicable die area (mm²)
# Figures: dense peaks from each vendor's official datasheet/product page (sparse ÷ 2)

# name: ([(dense TFLOPS/TOPS, bit length), ...], die area mm², total I/O GB/s, DRAM GB/s, datacenter part?)
CHIPS = {
    "A100 SXM":         ([(312, 16), (624, 8)], 826, 600 + 64, 2039, True),
    "A800 40GB Active": ([(623.8 / 2, 16), (1247 / 2, 8)], None, 400 + 64, 1555, False),
    "H100 SXM":         ([(989.4, 16), (1978.9, 8)], 814, 900 + 128, 3352, True),
    "H800 PCIe":        ([(3026 / 2, 8)], None, 400 + 128, 2000, True),
    "H200 SXM":         ([(1979 / 2, 16), (3958 / 2, 8)], None, 900 + 128, 4800, True),
    "RTX 4090":         ([(660.6, 8)], 608.5, 64, None, False),
    "L40":              ([(181, 16), (362, 8)], 608.5, None, 864, True),
    "L4":               ([(121, 16), (242, 8)], None, None, 300, True),
    "MI325X":           ([(1307.4, 16), (2614.9, 8)], None, None, 6000, True),
    "B200 (HGX)":       ([(2250, 16), (4500, 8), (9000, 4)], None, 1800 + 128, 7700, True),
    "B300 (HGX)":       ([(4500, 8), (13500, 4)], None, None, 7700, True),
}


def tpp(rates):
    # Take the largest TPP across bit lengths (Technical Note 2, item 2)
    return max(r * bits for r, bits in rates)


def rule_2023(t, pd, datacenter):
    # 3A090.a / .b as amended in October 2023 (same as the current text in November 2026)
    if pd is None:
        return "3A090.a (a.1)" if t >= 4800 else "PD needed"
    if t >= 4800 or (t >= 1600 and pd >= 5.92):
        cls = "3A090.a"
    elif (2400 <= t < 4800 and 1.6 <= pd < 5.92) or (t >= 1600 and 3.2 <= pd < 5.92):
        cls = "3A090.b"
    else:
        return "not 3A090"
    if not datacenter and t < 4800:
        return "excluded (Note 2)"
    return cls


def rule_2022(t, io):
    # October 2022 3A090.a: aggregate bidirectional transfer >= 600 GB/s AND TPP (then bit×TOPS) >= 4800
    if io is None:
        return "-"
    return "3A090.a" if (t >= 4800 and io >= 600) else "not 3A090.a"


def under_2026_limits(t, dram):
    # Numeric conditions of the January 2026 rule (91 FR 1684): TPP < 21,000 and DRAM bandwidth < 6,500 GB/s
    if dram is None:
        return "-"
    return "yes" if (t < 21000 and dram < 6500) else "no"


print(f"{'chip':17} {'TPP':>7} {'PD':>6}  {'2022':12} {'2023':18} {'2026 limits':>11}")
for name, (rates, die, io, dram, dc) in CHIPS.items():
    t = tpp(rates)
    pd = t / die if die else None
    pd_s = f"{pd:6.2f}" if pd else "     -"
    print(f"{name:17} {t:7,.0f} {pd_s}  {rule_2022(t, io):12} "
          f"{rule_2023(t, pd, dc):18} {under_2026_limits(t, dram):>11}")

# The L4's die area is not in the official table, so solve for the area boundaries where its class changes
t_l4 = tpp(CHIPS["L4"][0])
print(f"\nL4 TPP={t_l4:,.0f}: 3A090.a.2 if die <= {t_l4 / 5.92:.0f} mm2, "
      f"3A090.b.2 if {t_l4 / 5.92:.0f} < die <= {t_l4 / 3.2:.0f} mm2")
print(f"MI325X margin to 21,000: {21000 - tpp(CHIPS['MI325X'][0]):.1f}")
```

Output (pasted as is):

```text
chip                  TPP     PD  2022         2023               2026 limits
A100 SXM            4,992   6.04  3A090.a      3A090.a                    yes
A800 40GB Active    4,990      -  not 3A090.a  3A090.a (a.1)              yes
H100 SXM           15,831  19.45  3A090.a      3A090.a                    yes
H800 PCIe          12,104      -  not 3A090.a  3A090.a (a.1)              yes
H200 SXM           15,832      -  3A090.a      3A090.a (a.1)              yes
RTX 4090            5,285   8.68  not 3A090.a  3A090.a                      -
L40                 2,896   4.76  -            3A090.b                    yes
L4                  1,936      -  -            PD needed                  yes
MI325X             20,919      -  -            3A090.a (a.1)              yes
B200 (HGX)         36,000      -  3A090.a      3A090.a (a.1)               no
B300 (HGX)         54,000      -  -            3A090.a (a.1)               no

L4 TPP=1,936: 3A090.a.2 if die <= 327 mm2, 3A090.b.2 if 327 < die <= 605 mm2
MI325X margin to 21,000: 80.8
```

The "2026 limits" column only shows whether a chip meets the **numeric conditions** of the January 2026 rule (TPP below 21,000 and DRAM bandwidth below 6,500 GB/s). Getting case-by-case review also requires every certification in Section 5.6.

### 6.3 What the results show

**The A800 and H800 slipped through the 2022 "and," then got caught in 2023.** The A800 40GB Active has a TPP of 4,990, above 4,800, but its NVLink runs at 400 GB/s (bidirectional). The H800 PCIe has a TPP of 12,104, but its NVLink is also 400 GB/s. Lenovo's product guide says the differences between the H800 and H100 are "the NVLink interface speed and the FP64 performance." Because the 2022 text required a transfer rate of 600 GB/s or more **at the same time**, by my calculation these chips did not fall under the 2022 3A090.a. The 2023 amendment removed the transfer-rate condition from the item definition, and TPP alone put them in 3A090.a. NVIDIA's 10-K also lists the A800 and H800 among products affected by the October 2023 rule.

**The RTX 4090 gaming card also falls under 3A090.a.** Its TPP is 5,285, above 4,800. Because it is not a datacenter part, it might look eligible for the Note 2 exclusion, but that exclusion only applies below a TPP of 4,800. NVIDIA's filings include the RTX 4090 among products affected by the October 2023 rule. In this case the text points to License Exceptions NAC and ACA as possible paths.

**The L40 is a textbook 3A090.b case.** Its TPP is 2,896, short of 4,800, but its performance density of 4.76 puts it in 3A090.b. It shows how performance density works: a smaller chip is still controlled if it packs a lot of performance per unit area.

**The L4 cannot be classified without its die area.** Its TPP is 1,936. Working back from the density thresholds, it would be 3A090.a.2 with a die of 327 mm² or less, and 3A090.b.2 between 327 and 605 mm². NVIDIA's 10-K lists the L4 among affected products, but the Ada whitepaper's L4 table gives no die area. When a single number flips the class, check the manufacturer's classification rather than deciding on your own.

**The 21,000 line sits just above the MI325X.** The H100 and H200 both come out at about 15,830, well below 21,000 (the H200's memory bandwidth rises to 4.8 TB/s, but its TPP is almost the same as the H100's). AMD's MI325X comes out at 20,919, only 80.8 below the ceiling, and its 6 TB/s of DRAM bandwidth is under the 6,500 GB/s limit. The line sits right where the two chips the rule names, the H200 and the MI325X, just fit. That is how my calculation reads; BIS has not explained how it chose 21,000. The B200 (36,000) and B300 (54,000) exceed both the TPP and DRAM bandwidth limits.

**The highest-TPP rule bites at FP4.** The B300 does 4.5 dense PFLOPS at FP8 for a TPP of 36,000, but 13.5 dense PFLOPS at FP4 for a TPP of 54,000. The 13.5 is the eight-GPU total of "108 PFLOPS" (dense) on NVIDIA's HGX page and in the Blackwell Ultra datasheet, divided by 8. The same datasheet's per-GPU row says "14 PFLOPS," which would give 56,000. The conclusion is the same either way, but it shows that numbers can disagree even within one vendor's documents. FP4 throughput is not always exactly twice FP8, so the rule "compute for each bit length and take the highest" can make chips that are strong at low precision score higher.

**A reference point for a Chinese chip.** In a September 2025 keynote, Huawei said its Ascend 950, then in development, would deliver "1 PFLOPS in FP8." If that were a dense figure, the TPP would be 8,000. But this is an announced plan, not a product datasheet, and it does not say whether the figure is dense.

### 6.4 Limits of this calculation

- Whether a chip is "designed or marketed for use in datacenters" is a judgment, not arithmetic.
- For multi-die chips such as the B200, vendors do not publish die areas, so performance density cannot be computed.
- When a vendor's materials show several values, the text says to use "the highest value the manufacturer claims." Results can change depending on which document you use.
- Besides the published rules, some products are restricted by individual notices, as with the H20. A TPP below the thresholds does not guarantee a chip is uncontrolled.

## 7. China-specific products and NVIDIA's and AMD's filings

### 7.1 Reading fiscal years

NVIDIA's fiscal year ends in late January. Fiscal 2026 ended on January 25, 2026, and the second quarter of fiscal 2027 ended on July 26, 2026. Fiscal years run about a year ahead of the calendar label, so read the numbers carefully.

### 7.2 NVIDIA's filings

Here are the main export control statements in NVIDIA's filings, in time order.

| Period | What NVIDIA disclosed | Source |
|---|---|---|
| August 2022 | US restrictions affected the A100 and H100 and systems incorporating them | 10-K, 10-Q |
| October 2023 | New license requirements covered the A100, A800, H100, H800, L4, L40, L40S, RTX 4090, GB200 NVL72, B200 and others | 10-K, 10-Q |
| April 2025 | License requirement for the H20 and products with equivalent memory and interconnect bandwidth; USD 4.5 billion charge in Q1 of fiscal 2026 | 10-K, 10-Q |
| August 2025 onward | Licenses for some H20 products. The 10-K reports about USD 60 million in H20 revenue under those licenses; the 10-Q says Chinese government restrictions meant "we were unable to sell our H20 inventory" | 10-K, 10-Q |
| February 2026 onward | Licenses to ship small amounts of H200 to specific Chinese customers | 10-K, 10-Q |
| First half of fiscal 2027 | USD 0.4 billion charge for H200 excess inventory and related items; shipments under 1% of Data Center revenue in the latest quarter | 10-Q |

On the H20 charge, the 10-K says NVIDIA booked USD 4.5 billion in Q1 of fiscal 2026, and explains what it was for and why (quoted without the amount):

> … associated with H20 for excess inventory and purchase obligations, as the demand for H20 products diminished.

On the August 2025 licenses, the 10-K also mentions the government's share:

> USG officials expressed an expectation that the USG will receive 15% or more of the revenue generated from licensed sales of our products, but the USG did not publish a regulation codifying such requirement.

You may see reports that "the government takes 15% or more of sales." NVIDIA itself says officials "expressed an expectation" and states that **no regulation codified it**.

For the H200, the January 2026 rule and the presidential proclamation interact in an unexpected way. From the 10-Q:

> The licenses require that the H200s go through an inspection process in the United States prior to any shipment to the customer. As a result, any H200s shipped under the new licensing program are subject to a 25% tariff upon importation into the United States.

Because the rule requires third-party testing in the US, the chips must first be imported into the US, which triggers the proclamation's 25% tariff. NVIDIA adds that it has not been able to pass the tariff on to customers.

The 10-Q also addresses the Chinese government:

> The Chinese government has encouraged customers to purchase from our China-based competitors and discouraged customers from purchasing, importing, or using our data center products, including any China-specific product designed to comply with U.S. export controls.

Even when the US grants licenses, sales will not happen if China discourages purchases. NVIDIA gives this as the reason H200 sales have lagged despite the licenses.

In the 10-Q's geographic breakdown, "China (including Hong Kong)" was USD 7.88 billion in Q2 of fiscal 2027. But this is grouped by **the headquarters location of direct customers**, and the filing notes that the end customer and shipping location may differ. Since NVIDIA itself says it has been effectively foreclosed from the data center market, this figure should not be read as China data center sales.

### 7.3 AMD's filing

AMD disclosed a similar license requirement in an 8-K filed on April 16, 2025:

> The Export Control applies to the Company's MI308 products. The Company expects to apply for licenses but there is no assurance that licenses will be granted.

In the same 8-K, AMD said charges for inventory, purchase commitments and related reserves could reach up to approximately USD 800 million. That was an estimate at the time of filing, not a final figure.

### 7.4 Separating company claims from policy assessment

In its 10-Q, NVIDIA argues that the controls have helped Chinese competitors build their ecosystems and harmed its competitiveness. BIS, in its rules, says the aim is to keep advanced compute from being used for military and intelligence purposes. Both are claims by interested parties. In my view, the effect of the controls can only be measured against a world without them, and public information alone cannot settle the question. This article takes neither side.

## 8. China's technical challenges: EUV, yield and HBM

Primary sources on the state of China's chip industry are limited. Here I separate confirmed facts from estimates.

**Huawei's own view (a company statement).** In his keynote at Huawei Connect in September 2025, Huawei's Eric Xu said:

> The Chinese mainland will lag behind in semiconductor manufacturing process nodes for a relatively long time. Sustainable computing power can only be achieved with process nodes that are practically available.

The company's strategy is to accept the process lag and make up compute with "SuperPoDs" that gang together many chips. The same keynote laid out plans for the Ascend 950 series with Huawei's own HBM (HiBL 1.0 and HiZQ 2.0). The 950PR was slated for the first quarter of 2026, and HiZQ 2.0 was described as 144 GB with 4 TB/s of bandwidth. **As of November 17, 2026, I could not confirm whether these shipped on schedule.**

**EUV and lithography (what the rule text tells us).** As Section 2 showed, lithography tools below 193 nm (EUV) need a license for China. Without EUV, fabs have to shrink features with multiple passes of DUV (deep ultraviolet) lithography, known as multi-patterning. The extra steps are generally thought to hurt cost and yield. I found no primary source giving concrete yields for Chinese production lines, however. Reporters and research firms have published estimates, but I cannot verify them, so this article gives no numbers.

**HBM and component sourcing (reporting).** Bloomberg Law reported in October 2025 that the research firm TechInsights tore down Ascend 910C samples and found TSMC-made dies and HBM2E from Samsung and SK hynix. TSMC reportedly said the dies predated its 2020 halt of supply to Huawei. I have not read the TechInsights report itself.

**BIS's view (an agency claim).** In its May 13, 2025 guidance, BIS listed the Huawei Ascend 910B, 910C and 910D (printed as "Ascent 910D" in the guidance) as examples of chips presumed subject to GP10 restrictions, saying they were "likely developed or produced in violation of U.S. export controls." That is BIS's position, not a fact established in court or elsewhere.

This is also an example of wording broader than the rule. One bullet point in BIS's press release the same day said, in effect, that using Huawei Ascend chips anywhere in the world violates US export controls. But the guidance itself says such use "risks violating" the EAR and that a "presumption" applies. Do not mistake a press release's wording for the rule.

## 9. Implications for Japanese companies: Japan's FEFTA and the extraterritorial reach of the EAR

### 9.1 Japan's own controls (Foreign Exchange and Foreign Trade Act)

Japan runs its own export controls under the Foreign Exchange and Foreign Trade Act (FEFTA). These are the main semiconductor-related amendments I confirmed in Ministry of Economy, Trade and Industry (METI) documents and in the amendment history on e-Gov, Japan's official law database:

| Amendment | Promulgated | In force | Content (key points) |
|---|---|---|---|
| Amendment to the ministerial ordinance on goods | 2023-05-23 | 2023-07-23 | Added 23 types of chipmaking equipment (3 cleaning, 11 deposition, 1 annealing, 4 lithography, 3 etching, 1 inspection) |
| Amendment to the ministerial ordinance on goods | 2024-07-08 | 2024-09-08 | Added CMOS integrated circuits, scanning electron microscopes for chip imaging, quantum computers, GAAFET-related technology and more |
| Amendments to the Foreign Exchange Order, Export Trade Control Order and ministerial ordinance | 2025-04-03 (cabinet order on 2025-03-28) | 2025-05-28 | Added "advanced IC chips" and computers containing them. Also added equipment for upgrading lithography tools, EUV masks and more |
| Review of catch-all (supplementary) controls | 2025-04-09 | 2025-10-09 | For general destinations, exporters must apply for a permit when they judge there is a high risk of use in developing conventional weapons. Even for Group A countries, METI can notify exporters when diversion is a concern |
| Amendment to the ministerial ordinance on goods (METI Ordinance No. 72 of 2025) | 2025-11-14 | 2026-02-14 (some parts 2025-11-15) | Removed the 600 GB/s transfer-rate condition from "advanced IC chips" (Art. 6(1)(yo)) |

According to METI, list controls (which name specific items) apply to "all regions," and catch-all controls apply to "regions other than Group A."

The conditions for "advanced IC chips" (Article 6, item 1, sub-item "yo" of the ministerial ordinance) resemble US 3A090 but are not the same. When they took effect on May 28, 2025, they covered ICs with an aggregate bidirectional transfer rate over all inputs and outputs to other ICs (excluding volatile memory) of 600 GB/s or more, and with units whose TPP is 6,000 or more. An amendment in force from February 14, 2026 removed the transfer-rate condition. The current text on e-Gov reads only: ICs "having one or more digital processor units, with a total processing performance (TPP) of 6,000 or more."

| Item | US 3A090 (current) | Japan, Art. 6(1)(yo) (current, in force Feb 14, 2026) |
|---|---|---|
| Compute threshold | TPP 4,800 (.a), combined with performance density | TPP 6,000 |
| Transfer rate condition | None (removed from the item definition in 2023) | None (removed by the amendment in force Feb 14, 2026) |

Like the US, Japan has dropped the transfer-rate condition. Applying my calculation mechanically, both the H100 SXM (TPP 15,831) and the H800 PCIe (TPP 12,104) meet the TPP ≥ 6,000 condition, while the A100 SXM (4,992) and RTX 4090 (5,285) fall short of 6,000.

The detailed TPP rules (how to count bit length, how to treat sparsity) are not in the ordinance itself but in METI's operational circular. In METI's May 2025 explanatory material, the circular defined bit length as the largest bit length of the operation's "inputs or outputs," which differs from the US wording ("inputs to the multiply"). **I could not confirm the current circular after the February 2026 amendment as of November 17, 2026.** For a real classification, check the current circular and consult METI or an expert.

### 9.2 When the US EAR reaches Japanese companies

The EAR applies outside the United States when US-origin items, or certain foreign items, are reexported or transferred. These are the main situations Japanese companies should check:

1. **Dealing with companies headquartered in D:5 or Macau.** In its May 31, 2026 guidance, BIS stated that advanced items such as 3A090.a for those companies need a license **at every destination outside the United States**. Selling or transferring NVIDIA GPU servers to a Chinese-owned site in Japan can be covered. The guidance also says bona fide data center operators need not stop ongoing use of existing equipment "until further notice."
2. **Data center and cloud operators.** §744.24, scheduled to take effect on November 19, controls exports for data centers, wherever located, that provide large amounts of compute to companies headquartered in D:5 or Macau. Data centers in Japan are no exception. Once it takes effect, checking customers' headquarters, ultimate parents and use cases becomes important.
3. **Equipment makers.** The December 2024 rule created a list of countries excluded from certain chipmaking equipment license requirements (Supplement No. 4 to CFR Part 742). The list includes Japan and the Netherlands. Japanese equipment makers need to check both Japan's FEFTA controls and the US rules.
4. **The AI Diffusion Rule's effect is set to disappear.** Japan was among the 18 favored countries under the AI Diffusion Rule, but the rule itself is to be formally rescinded on November 19. Japan is not among the five countries (Indonesia, Malaysia, Singapore, Brazil, Thailand) that the new rule adds to the license requirement.

### 9.3 A practical checklist (for reference)

These are not legal conclusions. They are items worth organizing before you talk to an expert:

- The ECCN of each product (have you obtained the manufacturer's classification?)
- Not only the counterparty's location but also its headquarters and ultimate parent
- If you provide compute through remote access, who the users are and at what scale
- The classification under Japan's FEFTA (list controls and catch-all)
- Tracking amendment and effective dates (especially the November 19 rule)

## 10. What I could not confirm as of November 17, 2026

To be transparent, here is what I could not confirm.

1. The **published text** of FR Doc 2026-23529. This article relies on the public inspection copy and the pre-publication copy BIS posted (RIN 0694-AK50). Details may change in the published version.
2. NVIDIA's results and disclosures for Q3 of fiscal 2027. Anything released after this article was written is not reflected.
3. Official NVIDIA specs for the H20 and later China-specific products. I found no official figures, so I did not compute their TPP.
4. Whether the "15%" share tied to the August 2025 H20 licenses was later written into a rule. As of NVIDIA's 10-K, it had not been.
5. Whether the GAIN AI Act or the Remote Access Security Act (RASA), both mentioned in NVIDIA's 10-Q, became law.
6. Primary sources on yields at Chinese fabs, HBM output or EUV development.
7. Actual shipments of Huawei's Ascend 950 series.
8. Each product listed in the annex to Proclamation 11002.
9. The original text of Dutch government measures and ASML's individual statements.
10. The current METI operational circular for "advanced IC chips" (the TPP calculation rules). Also the content of two items in e-Gov's amendment history: the June 5, 2026 cabinet order consolidating provisions after a FEFTA amendment, and the amendment to the Export Trade Control Order promulgated October 9, 2026 and in force October 16, 2026.

## 11. Summary: how to read export control news

Finally, here is what to check when you read export control news.

1. **"Ban" or "license requirement"?** If it is a license requirement, check the review policy too (presumption of denial, case-by-case, presumption of approval).
2. **Publication date, effective date and compliance date.** It is common for all three to differ. Some rules are announced but not yet in force (the November 19, 2026 rule is one).
3. **Rule, individual notice, guidance or press release?** The H20 restriction was an individual notice, and the 15% share is not in any rule. Press release headlines can be broader than the rule text.
4. **Is the control keyed to the item, the counterparty or the use?** A chip below the TPP thresholds can still be restricted because of the counterparty's headquarters or the end use.
5. **Are the numbers dense, and the maximum over bit lengths?** Computing from sparsity figures doubles the TPP estimate.

US controls changed direction many times in the four years from 2022. In November 2026 alone came an extension of the Affiliates Rule's stay and the remote access rule. In this field, an explainer from six months ago can already be out of date. Read this article as a snapshot as of November 17, 2026. For any real transaction, always check the latest rule text and consult an expert.

For the change in NVIDIA's leadership in November 2026, see [Jensen Huang's resignation as NVIDIA CEO](/blog/jensen-huang-resignation-nvidia-ceo-transition-governance), which separates confirmed facts from reporting.

How the controls feed into NVIDIA's competitive position is covered, with scenarios, in [NVIDIA's moat is more than a CUDA monopoly](/blog/nvidia-competitive-advantage-cuda-ecosystem-future-scenarios).
