GAFAM NewsEssay
Isn't AI development really just about money? The winner-takes-all structure and where nations and companies stand
Why AI spending runs so high: the winner-takes-all logic behind it, and how the US, China, Europe, Japan and Big Tech compare.
- First published in Japanese
- English edition
Good evening. The Owl here.
If you follow news about AI development, you keep running into astronomical numbers. To be honest, the amounts have often been so huge that they didn’t feel real to me, and I would just watch them go by, thinking, “Here comes another number with way too many zeros.”
But when you think about it, that money isn’t simply a gamble. It isn’t a crowd swept up in a fever, charging in blindly, and it isn’t blind faith that “this is revolutionary!” either. There has to be a rational reason behind it.
So this time, let’s look for the reasons hidden behind these massive investments, and think together about where the AI race stands now and where it’s heading.
Why does massive AI investment decide who wins the race?
Let’s start with the key question.
“Why is so much money being poured into AI?”
Most likely because the players themselves are convinced that massive investment is what decides who wins and who loses.
Of course, the sheer size of the AI market is pushing investment along. But that is only the backdrop.
The direct reason massive investment can’t be avoided, I think, lies in the structure of the AI race itself. Namely, a structure that tends to shift toward winner-takes-all for whoever gets ahead first, and a structure in which getting ahead requires massive investment.
Once you understand these two mechanisms, it becomes much clearer why investment has ballooned to this size. Let me lay out that structure below.
1. The structure that leads to “winner takes all”
Today’s AI market is often compared with the search engine market of the past. In search, there were plenty of competitors at first, yet in the end Google, backed by its technology, came to dominate almost the entire market.
In the AI market, too, the following characteristics make it easy for whoever gets ahead first to build a decisive advantage.
An early stage with lots of ground still to break
AI still has a lot of room to develop. Hype phrases like “college-student level” usually refer to only a tiny slice of what AI does. It’s not so different from saying “a calculator calculates at an expert level.” In an early environment like this, a company that builds a technical lead has a good chance of seizing the initiative in one stroke.
The high versatility of foundation models
The transformer, the mainstream technology of the current AI boom, was developed as a language model, yet it handles a wide range of applications such as translation, image generation and drafting meeting minutes. This one model can drive leaps in results across many downstream fields. Put the other way around, this same property means that the winner of the foundation model race can swallow and rewrite a great deal of downstream effort all at once.
An evolutionary cycle in which AI produces AI
Some AIs, like DeepMind’s AlphaGo Zero, keep learning and improving through self-play alone.[1] Newer techniques include the “Self-Taught Optimizer (STOP),” where AI writes code to improve itself,[2] and models like SEAL that keep learning from new information.[3] Thanks to evolution loops like these, an AI that has built a lead can easily widen the gap further by evolving itself.
2. The structure that makes massive investment indispensable
So why do companies trying to get ahead in this race have no choice but to pour in staggering sums? Let’s sort out the factors.
Hardware and power constraints
Modern AI development can’t move forward without computing infrastructure that needs tens of thousands of GPUs, large data centers and enormous amounts of electricity. How much money can go into building that infrastructure, and how fast, is said to determine the speed and quality of model development.[4]
On top of that, the “neural scaling laws,” which say performance keeps improving as you scale up a model’s parameter count, training data and compute, have been demonstrated again and again, for example in the paper by Kaplan and colleagues,[5] and they give investment a further push.
Lately, more people point out that bigger models don’t necessarily translate directly into better performance.[6] Still, the odds look very low that we’ll soon see a revolutionary way to improve AI other than scaling, a breakthrough that slashes costs, or a new kind of processing such as quantum computers. The direction of relying on the brute force of hardware and power doesn’t look likely to change for the time being.
The burden of human resources
AI development can’t do without experts for model design and tuning, data preprocessing, MLOps and more. Personnel costs are estimated to make up 29–49% of total development costs, so the burden is extremely heavy.[7]
Recently, the race to invest early in talent seems to have escalated. There was a lot of buzz around news that Meta offered staggering sums, pay packages in the hundreds of millions of dollars, to poach AI talent.[8]
——
To sum up: the reason we can argue that deep pockets are decisive in the AI race is that two structures overlap. (1) A market structure in which whoever gets ahead technically tends to win everything, and (2) the fact that getting ahead requires massive investment in computing infrastructure, human resources and model scale. I think it’s safe to say this picture won’t be shaken for at least the next five years or so.
How money shapes the international race
Next, let’s look at the actual numbers and the differences in strategy for AI investment at the level of nations.
1. The United States: far ahead in investment, with a rich ecosystem
First, the US. A dynamic country where the private sector keeps the AI revolution spinning. Private AI investment in 2024 was about $109.1 billion (about ¥16 trillion). That’s 12 times China’s figure.[9]
The government isn’t lagging either. Between 2019 and 2023, the US government is estimated to have put a cumulative $328 billion (about ¥48 trillion) into AI, overwhelming other countries. Looking at how it was used, the largest share went to “decision science,” followed by “computer vision,” “autonomy,” “machine learning” and “natural language processing (NLP).” The biggest spender was the US Department of Defense, which makes clear that AI is the most important technology from a national security standpoint as well.[10]
Then there’s the “Stargate Project.” Announced at the White House in January 2025, it’s a plan by OpenAI, SoftBank, Oracle and MGX, a giant project aiming for up to $500 billion (about ¥74 trillion) in AI infrastructure investment over the next four years. As a first phase, $100 billion (about ¥15 trillion) has already gone in, starting with a data center under construction in Abilene, Texas.[11]
This is proof that the US is serious about making “the decisive move in AI strategy.” In a structure where huge national resources and the private sector act as one, you can glimpse its will to take the leadership that holds the future of AI.
2. China: catching up with state backing
Next, China. Its private investment in 2024 was a modest one-twelfth or so of the US figure, about $9.3 billion (about ¥1.4 trillion). However, this excludes infrastructure investment by the big players, or includes it only partially.
To make up for that, the government put $133 billion (about ¥19 trillion) into AI-related areas in 2019–2023, and in 2025 launched a state-led public-private fund, actively channeling money as a national strategy.[12]
The Chinese government also laid the groundwork for AI development and data management at the national level early on, through laws such as the Cybersecurity Law (in force since 2017) and the Personal Information Protection Law (PIPL, in force since November 2021).[13] Compared with Western AI companies that get held up by privacy issues, Chinese companies are in an environment where they can develop with peace of mind “under the state’s protection.”
In addition, China ranks among the world leaders in the number of papers and patent filings.[14] It’s an investment style that symbolizes national strength: competing on “speed and volume over quality.”
3. Europe: pushing the InvestAI plan
Europe, meanwhile, is a bit more cautious. Private AI investment of about €560 million (about ¥92 billion) has been confirmed for the major EU countries combined (Germany, France, Sweden and Spain), and including the other member states, the EU as a whole is estimated at around €600–700 million (about ¥1–1.2 trillion).[15]
Europe does have some internationally crucial players, like Mistral, which takes on foundation models, and ASML, which has a global monopoly in chipmaking equipment. Even so, it lags behind the US and China.
But in 2025 it announced InvestAI, an AI investment initiative on the scale of €200 billion (about ¥33 trillion), and is starting to put real effort in, for example by building four “AI gigafactories” (large computing centers).[16]
4. Japan: strong in chips, dependent on foreign foundation AI
Japan has excellent researchers and companies with a strong presence in semiconductors, but homegrown foundation model development is nonexistent, or where it exists, stays at a very modest scale. For now, the mainstream approach is to use foreign APIs such as OpenAI’s, and the reality is that it’s hard to build competitiveness of its own.
Perhaps aiming to break out of that situation, Prime Minister Shigeru Ishiba announced last November (2024) a new support framework: more than ¥10 trillion (roughly $68 billion) in public support for AI and semiconductors by fiscal 2030 (April 2030–March 2031), designed to draw more than ¥50 trillion (roughly $340 billion) in combined public and private investment over the next 10 years.[17]
For Japanese readers, it may be a little disappointing that SoftBank’s big investment is going into America’s Stargate Project. But perhaps that is exactly Masayoshi Son’s judgment: in AI investment, what matters is betting on the winning horse.
Global trends seen through the size and style of major companies’ AI investment
Finally, let’s look at AI investment by major company.
A caveat first: companies define “where AI investment starts and ends” differently, and only a limited number of them disclose it at all.
So here, with some help from ChatGPT, I made it the top priority to compare things “as apples to apples (with the same yardstick) as possible.” For that reason, please read the numbers strictly as “rough guides for getting a sense of scale.”
Specifically, for giant companies like GAFAM, I organized the figures around capital expenditure (CapEx) on AI infrastructure (data centers, GPUs and so on). Startups like OpenAI and Anthropic, on the other hand, don’t disclose CapEx the way listed companies do, so for them I used funding commitments directly tied to AI (large capital raises, chip procurement, infrastructure plans and so on). For exchange rates I used approximations of $1 = ¥150 and €1 = ¥165, and rounded figures as appropriate.
Please also note that the companies covered are only some of those focusing on AI model development and AI infrastructure; this is not a complete list of every company.
| Company | Metric (as like-for-like as possible) | Amount (USD) | Approx. in yen | Period | Source |
|---|---|---|---|---|---|
| Amazon | Total CapEx outlook for 2025 | > $100B | > ¥15 trillion | 2025 | [18] |
| Alphabet (Google) | 2025 CapEx guidance | ~$85B | ~¥12.8 trillion | 2025 | [19] |
| Microsoft | 2025 CapEx outlook | ~$80B | ~¥12.0 trillion | 2025 | [20] |
| Meta | 2025 CapEx range | $66–72B | ¥9.9–10.8 trillion | 2025 | [21] |
| Apple | 2025 CapEx outlook (relatively small share for AI) | $13–14B | ¥2.0–2.1 trillion | 2025 | [22] |
| Alibaba | Three-year investment plan to strengthen AI/cloud | $53B (cumulative) | ¥8.0 trillion (multi-year cumulative) | 2025–2027 | [23] |
| Tencent | 2024 actual CapEx / plans to increase it in 2025 | $10.7B | ¥1.6 trillion | 2024 | [24] |
| Huawei | R&D spending (company-wide, including AI) | CNY 179.7 billion ≈ $25B | ¥3.7 trillion | 2024 | [25] |
| OpenAI | Cumulative investment etc. by Microsoft | ~$13B (cumulative) | ~¥2.0 trillion | 2019–2023 | [26] |
| OpenAI | AI chip order with Broadcom | ~$10B | ~¥1.5 trillion | 2025 | [27] |
| OpenAI | Own commitment to the “Stargate” mega AI infrastructure plan | $19B | ~¥2.9 trillion | Next several years | [28] |
| Anthropic | Large funding rounds such as Series F | $13B | ~¥2.0 trillion | 2025 | [29] |
| Anthropic | Strategic investment by Amazon | up to $4B (cumulative) | ~¥0.6 trillion | 2023–2024 | [30] |
| Anthropic | Convertible notes etc. from Google | up to $2B (cumulative) | ~¥0.3 trillion | 2023–2024 | [31] |
| NTT | Medium-term investment in growth areas (data centers / AI, etc.) | - | ¥8 trillion (multi-year cumulative) | 2023–2027 | [32] |
| Rapidus (JSAM participant) | Planned investment in an advanced-chip (2nm, etc.) JV | - | ¥270 billion | 2024– | [33] |
| Toyota × NTT | Joint investment framework for next-generation mobility, AI and 5G research | - | ¥500 billion | Through 2030 | [34] |
| Sakura Internet | Investment plan for GPU infrastructure for generative AI | - | ¥113 billion (multi-year cumulative) | 2023– | [35] |
| Sony | Semiconductor investment plan (strengthening CMOS sensors, etc.) | - | ¥1.6 trillion (multi-year cumulative) | 2021–2026 | [36] |
| CyberAgent | R&D budget for AI and blockchain | - | ¥10 billion | 2023 | [37] |
Estimated AI development investment by major companies.
Here are the main takeaways from this table.
America’s tech giants and the “brute force of scale”
Amazon, Google, Microsoft and Meta are each pouring around ¥10 trillion (roughly $68 billion) a year into AI and data centers. Amazon in particular is above ¥15 trillion (roughly $100 billion), like a national budget. This sustained, massive spending is what gives them an overwhelming edge in the foundation model race.
Apple’s relative “restraint”
Apple’s AI-related investment is about ¥2 trillion (roughly $14 billion), a head shorter than the rest of GAFAM. It suggests Apple puts more weight on using AI in its own devices and ecosystem than on foundation models.
China fights with the public and private sectors together
Alibaba has announced ¥8 trillion over three years and Tencent around ¥1.6 trillion a year, but in single-year amounts they fall short of the US. What stands out, though, is how they work hand in hand with government support and regulation, accelerating research and market rollout as one.
Startups, too, play in the “trillion-yen” league
Even though OpenAI and Anthropic aren’t listed, they’ve secured trillion-yen-scale resources through outside fundraising and big chip procurement deals. With money and partnerships, they are closing in on the scale of the big companies.
Japanese companies are “off by an order of magnitude”
NTT expects ¥8 trillion (roughly $54 billion) in its medium-term plan, but that’s a five-year total, barely reaching a single year’s amount for the big US players. Sakura Internet, Sony, CyberAgent and others are smaller still, and the reality is that they’re allocating resources to applications and strengthening infrastructure.
Wrapping up
The AI race isn’t decided by technology alone. “How much money you can pour in” has become what separates winners from losers. Of course, the size of an investment doesn’t translate directly into results, but when the difference is a whole order of magnitude, it’s fair to say the gap can’t be ignored.
Looking at where things stand, multi-year commitments at the national level have grown to the scale of “tens of trillions of yen,” while at the company level the dividing line seems to be whether you can keep investing “trillions of yen” every year.
America runs out in front with overwhelming financial power, and China chases with the state and companies acting as one. Japan, meanwhile, isn’t joining the foundation model race and is trying to show a presence of its own in niche areas and in regulation and applications. That’s the picture that comes into clear view.
As for me, rather than being overwhelmed by investment figures in a league of their own, I’d like to quietly keep watch over the money flowing behind them and the ingenuity on the ground.
In the night forest, what truly rustles its wings is the “real” prey. Even on a night when a rain of trillions is falling, I want to keep ears that won’t miss a small drop of wisdom. 🦉
References
[1] Google DeepMind: AlphaGo Zero - Starting from scratch
[2] arxiv: Self-Taught Optimizer (STOP): Recursively Self-Improving Code Generation
[3] wired: This AI Model Never Stops Learning
[4] Morph Cast: AI Development Costs in 2024: Lower Your Bill with Easy-to-Integrate FER AI
[5] Cornell University: Scaling Laws for Neural Language Models
[6] Business Insider: Meta’s Yann LeCun: Scaling AI Won’t Make It Smarter
[7] arxiv: The rising costs of training frontier AI models
[9] HAI Stanford University: Artificial IntelligenceIndex Report 2025 - Economy
[10] Investopedia: Which Countries Are Investing Most in AI?
[11] TIME: What to Know About ‘Stargate,’ OpenAI’s New Venture Announced by President Trump
[12] Investopedia: Which Countries Are Investing Most in AI?
[13] wired: Ignore China’s New Data Privacy Law at Your Peril
[14] Science.org: China tops the world in artificial intelligence publications, database analysis reveals
[15] European Parliament: AI investment - EU and global indicators
[16] European Commission: EU launches InvestAI initiative to mobilise €200 billion
[17] Ministry of Economy, Trade and Industry: Press Conference by Minister Muto (Excerpt)
[18] DataCenterDynamics: Amazon 2025 capex to reach $100bn, AWS 2024 revenue hit $100bn - DCD
[19] Reuters: Google parent Alphabet surprises with capital spending boost after earnings beat
[20] Reuters: Microsoft plans to invest $80 billion on AI-enabled data centers in fiscal 2025
[21] Meta Reports Second Quarter 2025 Results
[23] Reuters: Alibaba to invest more than $52 billion in AI over next 3 years
[24] Reuters: Tencent joins China’s AI spending race with 2025 capex boost
[27] Reuters: OpenAI to launch its first AI chip in 2026 with Broadcom, FT reports
[28] Reuters: OpenAI, SoftBank each commit $19 bln to Stargate AI data center, the Information reports
[29] Wall Street Journal: Anthropic Valuation Hits $183 Billion in New $13 Billion Funding Round
[30] Amazon: Amazon completes $4B Anthropic investment to advance generative AI
[31] Reuters: Google agrees to invest up to $2 billion in OpenAI rival Anthropic
[33] Wikipedia: Rapidus
[35] Sakura Internet: Fiscal Year Ended March 2025 Financial Results