跪拜 Guibai
← Back to the summary

The AI Race Beyond the US and China: Europe Sets Rules, Gulf Buys Compute, Japan and Korea Build Factories

Core conclusion: Other countries have not exited the AI race; they have simply changed the event. In 2026, global AI compute will be about 45 GW: the US holds 35 GW (78%), China 5 GW (11%), and the EU about 2 GW (5%). On frontier models, nobody besides the two giants has a seat. So the rest competes on three things: the Gulf pays for compute (Saudi HUMAIN ~$100 billion, UAE Stargate 1 GW cluster), Europe trades rules for time (AI Act high-risk obligations pushed to 2027–2028, while building up to 7 superfactories), and Japan and Korea weld AI into factories (Japan METI 1.23 trillion yen for fiscal 2026, Korea ~550 trillion won for AI data centers). India and Singapore run a fourth route: application scale and governance hubs.

Table of Contents

1. One-sentence conclusion: not exiting, but switching tables

Let's get the answer first 🎯: "US-China AI race" only holds on two lines: frontier models and venture capital. Change the line and the table is crowded—the EU writes rules, the Gulf buys electricity and chips, Japan and Korea rebuild factories, India spreads applications, Singapore works as a governance hub.

The illusion that "only two players exist" comes from using model leaderboards as the only scoreboard. Leaderboards count only "who trained the strongest model," and that threshold is compute (算力) — in 2026, global AI compute is about 45 GW; the US alone holds 78%, China 11%, and the EU only 5%. With the leaderboard as the only lens, naturally you see only two players.

In my October 2026 AI roundup: world models, open-source trillion-parameter models, and agent battles, I summarized the month's international moves, and the most common comment was "how come it's all US and China stuff." This article swings the camera to the other side and looks specifically at what those other dozens of countries are doing, why, and whether their accounts add up.

2. Read the cards first: the three lines of 78% / 11% / 5%

Split "AI strength" into three quantifiable lines, and you see a very counterintuitive picture: the #1 country on each line is not the same one.

Dimension #1 #2 #3 Data source and scope
AI compute installed US 35 GW (78%) China 5 GW (11%) EU 2 GW (5%) Bruegel 2026, global ~45 GW
2025 important models US 59 China 35 Korea 8 Stanford AI Index 2026
AI paper citation share China 20.6% Europe 19.5% US 12.6% AI Index 2026 (2024 data)
2025 private AI investment US $285.9B China $12.4B — AI Index 2026, global $344.7B

Reading these three lines, the conclusion is clear 🔍:

There is also a statistical-scope issue that is easy to miss ⚠️: private investment of $285.9B vs $12.4B looks like a 23.1x gap, but the AI Index adds a footnote: between 2000 and 2023, China deployed roughly $912B through government guidance funds, of which an estimated $184B went directly into AI. Comparing the "private VC" column against "national mobilization" is not comparing the same pool. That principle appears throughout this article: check the scope before the multiplier.

References:

3. Why share is crueler than absolute: the denominator also grows

Most analyses skip one detail: a share is a fraction, and the denominator itself is growing.

Bruegel estimates global AI compute will expand more than 8x by 2031; the Swiss AI Institute (SIAI) further calculates that even if Europe completes all its underway projects and grows its own compute 10x, its global share climbs from 5% to only 5.6%—the absolute gap keeps widening.

An analogy 🎭: this is like the class average rising every year. You improve from 60 to 70, but the average moves from 65 to 85, so your rank still falls. If your catch-up speed trails the general speed, your rank never catches up.

Let's make the arithmetic concrete (Section 10 has runnable code to recalculate):

This is the first counterintuitive trade-off in this article: if share is bound to be missed, why does Europe spend up to €10 billion in public funds on gigafactories? The reason isn't "win training" but "win inference and data residency"—detailed in §4 (Europe) and §8 (ceiling). Simply put: training can be rented from others, but banks', hospitals', and governments' inference requests cannot leave national borders.

References:

4. Europe's route: rules for time

Europe plays a "three-piece" combo: rules first, compute second, company fallback.

First, rules. After the AI Act took effect in 2024, implementation lagged—standards for high-risk systems, member-state supervisory authorities, and conformity assessment bodies were not in place. So in November 2025 the European Commission proposed the Digital Omnibus on AI; the European Parliament passed it in June 2026, explicitly postponing high-risk obligations:

The essence is "time for standards": lock in the rule framework, get global vendors to build products around European compliance documents, risk assessments, and technical documentation formats, and only start issuing fines once standards and institutions are ready. For European domestic companies, this is a precious breathing window; for US and Chinese companies, it is a mandatory "compliance tax."

Second, compute. In July 2026 the EU launched a tender to build up to 7 AI gigafactories, each with at least 100,000 cutting-edge AI chips; public funding tops out at €10 billion, with an official expectation of leveraging at least €20 billion in private investment, for a combined €30+ billion. Bids close December 3, with results expected in early 2027. That same month, Nokia, Nscale, Telia, Verda, Skeleton, and others formed a "Northern AI gigafactory" alliance, building distributed compute across Finland, Estonia, Latvia, and Sweden, stitching the four small countries into one big network—if one point can't win, stitch four small countries into one.

Third, company. In September 2026, Paris-based Mistral raised a €3 billion Series D led by Samsung Electronics, at a post-money valuation above €21 billion; it also released a preview of Mistral Large 4 with a trillion parameters and started building a data center south of Paris targeting 1 GW. Europe finally has one company that can sit at the "frontier model" table—even if it is just one.

So Europe's real strategy: not competing on whose model is strongest, but on whose model is most compliant and easiest to get into European banks and government departments. It sells market access, not benchmarks (评测榜单).

References:

5. The Gulf's route: oil dollars buy a "compute landlord" seat

Saudi Arabia and the UAE take another path: don't train models yourself, buy your way to being a compute landlord.

Saudi Arabia: HUMAIN, wholly owned by sovereign fund PIF, chaired personally by the crown prince, spans data centers, chips, cloud, Arabic foundation models, and applications. It has committed roughly $100 billion, laying out 11 data centers with 2.2 GW capacity; its partnership with NVIDIA starts with an initial 18,000 GB300 Grace Blackwell systems, planning up to 500 MW of AI load.

UAE: Abu Dhabi's G42 leads Stargate UAE, building a 1 GW NVIDIA cluster with OpenAI and Oracle; the whole campus plans up to 5 GW, with the first 200 MW expected online this year—this is the largest such facility outside the US. On the capital side, MGX closed a $49 billion fund in June 2026, oversubscribed; its co-founder Mubadala manages $330 billion in assets and has listed AI and robotics as primary investment directions.

But a premise must be made explicit ⚠️: the Gulf's compute runs entirely on US-approved silicon. Chip export licenses sit in Washington's hands, so its "sovereignty" is conditional—data can stay onshore and models can be deployed locally, but the switch on compute is not in its own hands.

An analogy 🎭: this is like building your own power plant next to someone else's grid. You bought the plant, the land, and the generators, but the core components of the generator must be supplied by the other side, and a single licensing decision can decide whether you can expand.

This explains the Gulf's next move: buy chips on one side, be an LP on the other. MGX and PIF invest heavily in OpenAI, xAI, and the like, essentially trading capital for "priority model supply rights"—if you can't buy technology leadership, buy a shareholder seat in the company that has it.

References:

6. Japan & Korea: no frontier models, AI welded into factories

Japan and Korea make a more pragmatic choice: let others fight over general-purpose models; we do "physical AI" and upstream components.

Japan significantly increased spending in 2026. METI's fiscal 2026 support for AI and semiconductors is about ¥1.23 trillion, roughly 4x the previous level; ¥387.3 billion goes specifically to domestic foundation models, data infrastructure, physical AI, and underlying compute. The concrete projects are also very "Japanese":

Korea is even more forceful. In June 2026, a presidential "three super projects" plan named memory semiconductors, physical AI, and AI data centers as the three priorities, with total scale exceeding 1,800 trillion won; the AI data center piece alone is about 550 trillion won, targeting gigawatt-scale super-large facilities. The 2027 budget is the largest in history, with AI-related spending nearly doubling: 4.7 trillion won supports domestically developed frontier models, 500 billion won funds physical AI demonstrations across 8 sectors including manufacturing, smart factories, and defense; the National AI Computing Center lands in Haenam County, South Jeolla, built and operated by a consortium led by Samsung SDS.

Why do these two choose this route? Because their strongest asset is not a model team, but the supply chain: Japan's machine tools, materials, and cars; Korea's memory, displays, and batteries. For them, AI is not "a stronger chatbot" but "an extra sales cycle for hardware already sold worldwide." Model wins and losses are decided by factory orders, not leaderboards.

References:

7. India, Singapore, UK: application scale and governance hubs

The three routes above bet on compute or manufacturing. The remaining countries make a lighter bet: don't build the biggest machine room or the strongest model; use market size and rule-making voice for position.

India takes the "compute as public good" route. The IndiaAI Mission has a total envelope of ₹103.72 billion over 5 years, treating compute like tap water: the shared compute pool expanded to 45,000 GPUs by June 2026, with another 20,000 in procurement; the early 38,000 GPUs were billed at ₹65 per hour, about one-third the global average, with 1,050 TPUs as supplement. By August 2026, 237 projects had consumed 9.318 million GPU hours, and 12 teams made the shortlist for domestic foundation models.

Here is a directly convertible comparison ⚠️: one EU gigafactory's threshold is 100,000 cutting-edge AI chips; India's entire national public compute pool is 45,000 GPUs—less than half a factory. Not because India lacks money, but because it chose something else: build the developer base and usage habits first. In February 2026 it also held the India AI Impact Summit in New Delhi, with the same mindset as running an industry trade show—compute may be short, but voice must not be short.

Singapore takes the "governance hub" route. It invests S$1 billion in AI R&D from 2025 to 2030; in February 2026 it set up a National AI Council (NAIC) chaired directly by Prime Minister Lawrence Wong; in May it updated its National AI Strategy (NAIS) with 10 reprioritized directions, and established partnerships with Google and OpenAI, targeting 10,000 enterprises using AI within three years. Its size guarantees it can never win on compute, but on how rules are set and how cross-border data flows, it gets a seat at the table.

The UK sits between: it designated 5 AI Growth Zones (the first in Lanarkshire, £300 million, 3,400 jobs), committed £2 billion to scale public compute 20x by 2030, brought Bristol's Isambard-AI online, set aside up to £500 million for a sovereign AI unit to invest in domestic companies, and officially projects over £28.2 billion in investment by 2030.

The common thread is clear: they compete not on who has the strongest model, but on who can get AI adopted by their enterprises and governments fastest.

References:

8. The shared ceiling: electricity, chip permits, and talent

Whatever route a country takes, it eventually slams into the same wall, and the three bricks come in a fixed order of difficulty: electricity can be bought with money, chips depend on approvals, and talent is hardest to keep.

First brick: electricity. In 2026, global AI data-center electricity demand is 29.6 GW, roughly equal to New York State's peak load—the difference is only 1.9 GW. Gartner estimates global data centers use 565 TWh in 2026, up 26% year over year (2025: 447 TWh), breaking 1,200 TWh by 2030—more than Japan's entire annual consumption. Goldman Sachs is more concrete: by end of 2028, the US alone faces a 33 GW electricity gap, roughly equal to the UK's baseload.

Second brick: chips. Chip export licenses are not in the buyer's hands. No matter how much money the Gulf has, its expansion pace depends on Washington's approvals—which is exactly why it builds machine rooms and acts as an LP at the same time.

Third brick: talent. Stanford's AI Index 2026 contains a rarely quoted observation: America's ability to attract global AI talent is declining. This opens a window for other countries—Japan uses a $6.3 billion consortium to keep its engineers in its manufacturing system, Singapore uses governance and research positions to attract regional talent, and the logic is the same.

One sentence to sum up this wall: what money can solve is not a bottleneck; the bottleneck is exactly what money cannot solve.

References:

9. A comparison table: symptom → cause → fix

Compress the pitfalls from the previous eight sections into a table; when you hit a symptom, look it up directly:

Symptom Cause Fix
Feels like "only the US and China are doing AI" Only looking at the leaderboard as the one scoreboard Switch to three lines: compute installed, paper citations, private investment
Europe invests more and more, share keeps falling Global denominator expands 8x; growth cannot keep pace Give up share targets; pivot to inference and data residency
Gulf bought chips, expansion still depends on others Chip export licenses not in its own hands Build and invest in parallel; use LP seats to buy priority supply
Japan/Korea have no chance at general-purpose LLMs Lack frontier model teams and venture capital Move to physical AI and upstream components; score by factory orders
India/UK compute scale is an order of magnitude smaller Actively chose application and governance routes Use public compute pool to lower unit price; use rules to build ecosystem
All non-US-China projects block on the same spot Electricity, permits, talent are simultaneously constrained Lock electricity first, then chips, then talent

10. Fully runnable code: recalculate every country's share account

The code below depends on no third-party libraries; copy, paste, and run. It recalculates three things: current shares, how many GW are needed to hold a target share by 2031, and the EU superfactory's public-capital leverage ratio.

# -*- coding: utf-8 -*-
"""Recalculate this article's share math: current distribution, required compute for target share, public-fund leverage."""

GLOBAL_GW = 45.0                                                 # 2026 global AI compute (GW), Bruegel scope
CAPACITY = {"美国": 35.0, "中国": 5.0, "欧盟": 2.0, "其他": 3.0}   # region → installed capacity (GW)

def share_table(total_gw, capacity):
    """Print a regional compute share table.
    Args:
        total_gw: global compute total (GW)
        capacity: region → compute (GW) dict
    """
    for region, gw in capacity.items():                           # compute percentage per region
        print(f"{region}  {gw:>4.1f} GW  {gw / total_gw * 100:>5.1f}%")  # e.g., "欧盟   2.0 GW    4.4%"
    print(f"合计 {sum(capacity.values()):>4.1f} GW")               # verify sum of regions

print("=== 2026 年现状 ===")
share_table(GLOBAL_GW, CAPACITY)                                 # US 77.8% / China 11.1% / EU 4.4%

print("\n=== 全球扩容 8 倍后,守住目标份额需要多少 GW ===")
total_2031 = GLOBAL_GW * 8                                       # 2031 global total: 45 × 8 = 360 GW
for target in (0.05, 0.10, 0.15):                                # try 5% / 10% / 15%
    need = target * total_2031                                    # target share × global total = required GW
    multiple = need / CAPACITY["欧盟"]                            # relative to current EU 2 GW
    print(f"目标 {target:.0%}:需 {need:>5.1f} GW,是当前欧盟的 {multiple:>4.1f} 倍")

print("\n=== 两种口径下的中美差距 ===")
print(f"算力:{CAPACITY['美国'] / CAPACITY['中国']:.1f} 倍")        # 35 / 5 = 7.0x
print(f"私人 AI 投资:{285.9 / 12.4:.1f} 倍")                     # AI Index 2026: $285.9B vs $12.4B

print("\n=== 欧盟超级工厂的公共资金杠杆比 ===")
public, private = 10.0, 20.0                                      # public max €10B, expected leverage €20B
print(f"公共 {public:.0f}B → 私人 {private:.0f}B,杠杆比 {private / public:.1f}x")  # 2.0x

Actual output:

=== 2026 年现状 ===
美国  35.0 GW   77.8%
中国   5.0 GW   11.1%
欧盟   2.0 GW    4.4%
其他   3.0 GW    6.7%
合计 45.0 GW

=== 全球扩容 8 倍后,守住目标份额需要多少 GW ===
目标 5%:需  18.0 GW,是当前欧盟的  9.0 倍
目标 10%:需  36.0 GW,是当前欧盟的 18.0 倍
目标 15%:需  54.0 GW,是当前欧盟的 27.0 倍

=== 两种口径下的中美差距 ===
算力:7.0 倍
私人 AI 投资:23.1 倍

=== 欧盟超级工厂的公共资金杠杆比 ===
公共 10B → 私人 20B,杠杆比 2.0x

The key to reading this output is the second block: under the same "global expansion 8x" assumption, holding 5% requires growing 9x, pushing to 15% requires 27x. And the EU's 7 gigafactories under construction each have a 100,000-chip threshold; even if all are built, they fall far short of 27x. That is the "denominator grows" point from Section 3.

11. Three ways to bet: choose a route by the cards in your hand

Compress all the cases above into a decision table:

Thickest card in hand Route to take Representatives Key move
Cheap electricity, lots of money Buy compute and become a landlord Saudi Arabia, UAE Build GW-scale campuses; use LP seats for priority model supply
Industrial scenarios and supply chain Physical AI + upstream components Japan, Korea Wedge AI into machine tools, cars, memory, and displays
Market and rule-making power Compliance as moat EU Use AI Act to make technical document format its own standard
Population and delivery capability Public compute + application scale India, Singapore, UK Lower compute unit price; build developer and governance ecosystem

The usage is to check each item in order, and the order cannot be reversed: first ask whether you have electricity—if not, don't touch self-built machine rooms; then ask whether you have scenarios—if not, don't do vertical domain models; only then come "rules" and "ecosystem" as bonus items after the first two cards are played. Betting in reverse is buying the thing you are least good at with the most expensive method.

When to read this line: when a product goes overseas, builds compute abroad, or evaluates a country's compliance cost. When not: when developing only domestic applications—these geopolitical variables won't change your technical choices in the short term.

This column, LLM Large Model Technology Column, moves from the Transformer attention mechanism all the way to model selection, inference optimization, agent deployment, and AI industry structure. Each piece includes a runnable code block and updates weekly. To read in order, subscribe on the column page.

Comments

Top 1 of 2 from juejin.cn, machine-translated. The original thread is authoritative.

ET外星人 1 likes

Next, I'll tell you the conclusion in the most straightforward, no-nonsense way [polite smile]

ENCI

Haha, you saw through me 🤣 The title itself is the condensed conclusion—Europe's 5% of compute goes to compliance frameworks, Saudi Arabia splashes out on data centers, Japan and Korea go deep into vertical scenarios. The article is full of details, but the core judgment is just these three sentences~ Thanks for the follow-up comment!