*Yan X Zhang is an Associate Professor in the Mathematics and Statistics Department at San Jose State University, with degrees in mathematics from Harvard (BA) and MIT (PhD). He was one of the curriculum designers and instructors of UC Berkeley’s Stat 157 (*[*<u>stat157.com</u>*](http://stat157.com/)*), one of the first university classes on forecasting. His views in this piece are his own and do not reflect those of his employers.*
# **Project Overview**
In this piece, I want **to study the intersection of geopolitics and AI, pose mental models, and make predictions.**
This project is highly complex. Specifically:
1. **Scope:** "AI and geopolitics" touches too many topics for anyone to have complete expertise on the following. (And I'm not a domain expert in even one of these.)
* wars
* cyberwars
* LLMs (Large Language Models)
* how AI companies work; how governments interact with them
* energy
* AI security regulations (and whether they will accomplish anything)
* regulation relevant to geopolitics + AI
2. **Model complexity:** the number of moving parts is high. I can imagine a forecaster with a model where changing the probability that there is a DRAM (Dynamic Random-Access Memory) breakthrough of some particular type in the next 10 years from 50% to 60% might change her estimation of "who will win the AI race" significantly.
3. **Very different starting points:** Slight differences in worldviews can make a person reject a model or prediction for visceral reasons. For example, different answers to "do countries have long-term goals?" or "do countries mostly operate on realpolitik?" can lead to fundamentally incompatible assumptions. I have met very intelligent people who accepted some of my assumptions as "obvious" and equally intelligent people who thought the same assumptions were "obviously wrong."
My main reaction to this complexity was to focus on **what a reader like myself would want**. After two years of meditation on this problem and grant, including consulting \~20 people for both their knowledge and what they might want from such a project, I arrived at these goals:
1. To provide **"hard data" that would be useful for most people** and to pose **low-complexity models**. This way, readers with better domain knowledge or who prefer higher-complexity models can take what they want from this work and disregard the rest.
2. To make **falsifiable** predictions, but with **less than Metaculus main-feed question** rigor¹. This is a compromise between Metaculus precision (which people broadly prefer in AI discussions) and broad-stroke directional correctness (which people broadly prefer in geopolitical discussions).
3. To make all predictions for **December 31, 2030**². I will label more speculative longer-term predictions separately.
My outline is:
1. I define the **(major) players**³ to be the US and China⁴.
2. I define **resources** to be "**quantitative and monotonic**⁵" factors that correlate with "being more powerful in AI" that most forecasters and agents are likely to consider in their models. All forecasters care about capital, for instance, though with different weights and different causal chains. I divide the resources loosely into three types:
1. **Semiconductors:** the most important resource for AI (and top-tier importance even without it)b. **Fungible resources** such as **capital**, **energy**, or **raw materials**c. **Non-fungible resources**, focusing on **human capital** (skilled workers / AI researchers) and **data** (collection, organization). I also introduce the idea of **"intellectual market share"** to capture the nebulous idea of "where will the relevant human capital want to go."
3. I will then focus on one concept per section, mostly driven by what my discussion partners found the most interesting.
1. For the first few sections, I will summarize the data/intuitions I have regarding the resources above. I will make some predictions as well.
2. I will analyze Taiwan and TSMC (Taiwan Semiconductor Manufacturing Company).
3. I end with a batch of more speculative and less falsifiable predictions that I think are worth mentioning.
To limit scope-creep, I decided to **minimize Artificial General Intelligence (AGI) discussion**. This is obviously important—with short enough timelines, nothing in this document matters anyway! My intuition is that state actors will think of AGI as important but still hedge it with other aspects of AI like military and economic dominance. So it made sense for me (and many of my begrudging discussion partners) to model the states as playing a general "AI race" rather than be sniped by models of AGI phases and timelines. In other words, **I am trying to predict what state actors would do regarding AI before the point where AGI becomes the visibly dominant point of conversation**.
**Final disclaimer:** While I am using many words that suggest a zero-sum "race" between the major "players", I find this language distasteful and often counterproductive⁶. However, I am very sure that the agents I am modeling will think in this way, so using this language better reflects their intentions and thought processes.
# **Semiconductors**
The key resource relevant to AI is **semiconductors,** a cluster that includes chips, foundries, machines, and relevant science. Semiconductors are so ubiquitous they would be very highly valued as a strategic resource for states even not accounting for AI. They are akin to oil in the sense that state actors go to extremes to compete for them.
The semiconductor industry is highly globalized and stratified. There are many steps of processing—from materials to chip fabrication to chip design to manufacturing⁷—but at each step there tend to be just 1-3 key players.
At the step of chip manufacturing (the main output of the industry and consumed by AI companies), it is useful to think of 3 main types of chips: **logic chips, memory chips,** and **others**.
1. **Logic chips** do computation and are the most crucial for AI advancement (training, inference, etc.). The most important (high end) chips can only be manufactured by **TSMC** (clear lead), Samsung, or Intel.
2. **Memory chips** are less "fancy" than logic chips but are also important enough that they can become bottlenecks⁸. They're basically split into DRAM (3 dominant companies: SK Hynix \~35%, Samsung \~35%, Micron \~25%) and NAND (Not AND), which is dominated by Samsung at \~35% but flatter afterwards with SK Hynix at \~20% and several runner-ups around 10% each starting with Kioxia.
3. Other chips (mostly less powerful, also includes analog) are used everywhere and should have moderate economic / domestic political effects on their own, but chips used in AI will be dominated by logic and memory chips.
There are many other things to consider / learn about semiconductors. In the context of geopolitical discussions, I think the main additional factors to consider are:
1. **Export controls**: semiconductors are one of the most importantly controlled markets (evidence for state actors' perceived importance of them).
1. The most important export controls are US/Japanese/Dutch exports stopping high-end **SMEs (semiconductor manufacturing equipment)** and additional controls such as the Foreign Direct Product Rule that forbids classes of chips made involving American tooling to be sold to China at all.
2. As a first order, I consider export controls **fairly important**, and their unpredictable arrival is a source of a lot of noise. However, the market is somewhat homeostatic. For example, NVIDIA reacted to chip bans to China by degrading their chips to a level suitable for export controls, smuggling is rampant, and one can do AI training overseas to avoid export controls. So I tend to **mildly downgrade** the effects of export controls.
2. There are many other steps in the industry. Of them, probably the most important is **extreme ultraviolet lithography (EUVL / EUV)**, the "printing" process that puts a circuit on the wafers.
1. This is dominated by ASML (Advanced Semiconductor Materials Lithography), one of the EU's main "cards" in the geopolitical game. China cannot currently get any EUV machines due to export controls. However, ASML's importance factors through its consumers—primarily TSMC in our case—so the marginal value of considering it in our discussion seems low.
2. China has invested a lot into developing its own path (and/or alternatives) to EUV. While success would indeed drastically change the relevant importance of TSMC and other factors, I find evidence of China's progress to be hard to analyze and often overhyped. Thus, I will **mostly ignore** this possibility.
**At a Glance:** There are broadly three types of chips. I expect the players to grapple for mastery over all three types with tactics such as export controls. Given the current power balance, this will probably mean the US and its allies controlling China rather than the other way around. Even the less important types of chips could be held as leverage if there is de facto control. The different "layers" of the chip manufacturing process can also serve as individual points of contention, usually controlled by 1-3 parties, with the most "obviously valuable" piece being TSMC's control of logic chips.
Predictions:
1. For logic chips, the main battleground is clearly Taiwan/TSMC; I delay this discussion to the Taiwan section.
2. For memory chips, I expect the US to heavily support South Korea as an ally in the upcoming years (mostly due to Samsung and SK's memory chip capacity) and for China to take stabs at memory chips (CXMT, etc.).
1. As a proxy, I expect (75%) at least 1 event in the same reference class / scale of Tsinghua Unigroup trying to acquire Micron or a massive injection of resources into Fujian Jinhua by 2030.
3. More generally, I expect (90%) the US to introduce at least two new "noticeable escalations" of export controls (possibly through an ally) by 2030.
1. Things in this reference class would be the CHIPS Act or the most recent (6/2025) Taiwan putting Huawei on its entity list.
2. I would also include control through human capital; for example, a hypothetical escalation disallowing Chinese nationals / Chinese American citizens to return to China after working at a top US-based AI lab. See **Human Capital** later.
4. I expect China to also invest in trying to win the legacy chips battle (chips > 28nm) and use it as leverage⁹. I do not expect the US to put as much resources into these chips as logic and memory chips.
# **Fungible Resources**
While there are a ton of other resources to consider, after considering **capital**, **energy,** and **raw materials**, I personally believe marginal returns of studying a new resource drops off quickly. For example, industrial parks and industrial capacity can be bought or may be bottlenecked by a combination of capital, energy, and raw materials.
On capital:
1. We can roughly split the input into public (government) vs private (VC firms, etc.)
1. The main players use fairly distinct strategies. In the US, most of the innovation and money is in the private sector. In China, there is very tight integration between public/private sectors, and private sectors are often heavily subsidized or considered de facto extensions of the government.
2. The rest of the world tends to be somewhere in the middle. The EU has a less vibrant private sector and fairly heavy (but not as high as China) government support. Also see e.g. G42 from the UAE (United Arab Emirates), which has ties to the government and sovereign wealth funds.
2. **Public funding:** quick overview using public estimates:
1. The US: Total federal NITRD (Networking and Information Technology Research and Development)¹⁰ funding for the upcoming year will be 3.3 billion. For a different methodology, Stanford's AI Index report uses NLP (Natural Language Processing) to estimate about 4.4 billion in AI-related public grants. Note that this does not include one-time items such as the CHIPS act, which are injections on the orders of tens of billions USD/year but are more broad and include semiconductor industrialization.b. China: using similar methodology as CSET's (Center for Security and Emerging Technology) 2019 study, which takes as main input China's Ministry of Finance statistics, I crudely estimate¹¹ something between 3 and 7 billion USD. Also, see the "Big Fund"¹² currently with about 100B USD AUM (Assets Under Management) with a horizon for 2039.
3. **Private funding:** On paper, this is usually a much bigger source of money than public spending. For example, Project Stargate will aim for 500 billion dollars, with main equity backers being OpenAI, SoftBank, and Oracle. An estimate¹³ gives that the U.S. private AI investment hit \$109.1 billion in 2024 and China's at 9.3 billion.
4. The rest of the world:a. In the EU, Horizon Europe and Digital Europe are the main state funding sources for AI research¹⁴, accounting for a few billion USD/year. For private funding, see InvestAI's 200B Euro (of which 50B will come from public funding, in EU spirit) and European Innovation Council's 256M Euro.b. The Gulf states are also relevant, with Saudi Arabia and UAE (see e.g. Microsoft putting 1.5B USD into G42) as the biggest players. While they lack momentum, infrastructure, and human capital compared to the main players, they are putting in cash at a high rate.
**At a Glance:** my fairly consistent heuristic is that the US, China, and "the rest of the world" each put in comparable capital of a few billion USD/year into AI research, with the willingness to make specialized injections usually around a few tens of billions/year. Private funding is usually in the order of magnitude of hundreds of billions/year. I think only the orders of magnitude are important and I am fairly sure much of all players' accounting is purposefully obfuscated, such as by blurring the public/private lines.
On energy:
1. If we look at training costs¹⁵ from e.g. Epoch's breakdown¹⁶ of AI scaling constraints, we can see that energy is one of the main potential bottlenecks of scaling AI.
2. Production:
1. The dominant source of energy in the world is fossil fuels (about 80% of all production, split into oil, coal, and natural gas). The other sources are quite small, with nuclear (about 5%), and other types like hydro or biofuel shifting around 5-10%.b. For *electricity* in particular¹⁷, which is presumably more relevant for AI, the breakdown of energy types used are 35% coal, 22% gas, 15% hydro, 10% nuclear. Notably, only 3% oil!
3. For fossil fuels¹⁸, the three types are used roughly equally. For each resource:a. Coal: China at 50%, India and Indonesia both around 10%, Asia adding up to about 80% of all the world's production with other countries much smaller.b. Oil¹⁹: OPEC (Organization of the Petroleum Exporting Countries) accounts for 35%, extended parts of OPEC+ (mostly Russia) account for 15%. The US accounts for 20%. The trend is that non-OPEC is slowly rising in production compared to OPEC.c. Natural gas: US accounts for 25%, Russia accounts for 15%, the rest are much smaller (any individual country at most 5-6%)
**At a glance:** Energy is massively dominated by fossil fuels, though AI-relevant electrical energy downgrades the importance of oil. While I think energy is extremely important, I model the governments of the world as in an eternal race for energy. One could say the same about capital, but capital can be leveraged, credited, or delayed, whereas energy is needed to actually do things on time. Governments would **counterfactually be maximizing energy anyway**, worried about AI or not. Thus, in the context of predicting governmental actions specifically because of AI, energy is paradoxically **less** important. China is an energy powerhouse and can also more efficiently build and gather energy given state directives, though the US is also fairly self-sufficient.
On raw materials:
1. In my knowledge, the three most important raw materials for AI are:
1. Silicon: China makes 80% of the world's silicon production and the US around 10%. While the US does have a dependency on it, silicon is relatively easy to produce / acquire.
2. Germanium: China makes more than half of the world's germanium and banned it for US export in 2024. This is mostly in reaction to the 2022 US export controls.
3. Gallium: China produces more than 98% of the world's gallium. This was banned in conjunction with Germanium for US export.China has already escalated export controls²⁰ on many other raw materials, including antimony and graphite.
2. There are other materials (rare and non-rare) that are relevant to AI. For example, the US has recently added copper to its critical materials list (I assume this is relevant to AI through networks and data centers). I think what's more important here is not copper itself per se, but a signal that different bottlenecks are being hit as a result of high demand. I do not have a good strategy of how to incorporate these other materials for my predictions.
**At a glance**: China is the undisputed leader in raw materials, including all rare earth elements (REEs), and is likely to use this as leverage in the AI race. My assumption is that most of the relevant causal effects of raw materials factor through semiconductors. In other words, the main things we care about regarding AI and raw materials is how they would affect semiconductors, though they would of course have other effects on the GDP (Gross Domestic Product) through technology and manufacturing.
Predictions:
1. On capital, I expect mostly status quo:a. On public capital, I mostly (80%) expect the status quo of US and China both staying "on track" and putting somewhere around a total of 25B and 50B USD each into AI research during the 2025-2030 period. Keep in mind that much of the accounting could be invisible, obfuscated, or actively misleading.b. On private capital, I mostly (75% per year) expect the US to absorb between 100B and 150B of private investment a year, though I expect (60%) its fraction of all investment to **decrease** compared to China by 2030.
2. I expect the US to gravitate towards a tighter "integrated" model of public and private AI funding. This means some subset of:a. I expect 40% probability of the US labeling AI-related objects (such as best proprietary LLMs) objects of national security and subject to legal scrutiny by 2030.b. I expect 20% probability of the US visibly strong-arming AI companies to pool data / "work together" for national interest by 2030 (converging to 100% for long-term, though I am very unsure of timelines).
3. On minerals: here is one of the playing grounds where China is dominant, so I expect China to wield this advantage.a. I expect (70%) [<u>China to start / escalate controls on at least another valuable mineral from the US by 2030</u>](https://www.metaculus.com/questions/39282/new-prc-semiconductor-mineral-controls/).b. I expect a moderate chance (35%) of AI making a significant breakthrough in mineral identification / extraction such as when NETL (National Energy Technology Laboratory) identified²¹ an outsized mineral deposit.
4. On energy: I expect governments to mostly be maximizing energy anyway, but there are still some predictions:a. Because of the difficulty of oil being converted to electricity in particular, I expect OPEC States to be searching aggressively for alternate forms of energy, possibly trading oil at a discount for other energy forms.b. I expect a moderate chance (35%) of AI making a significant breakthrough²² in energy identification / conversion.
# **Non-Fungible Resources**
I limit my attention to **human capital** and **data**, though there are obviously many other angles (my instinct for the next three things to consider would be **intelligence / espionage**, **diplomacy**, and **robotics**).
## **Open vs Closed Models of Human Capital**
First, I wish to focus on the **AI researchers** (including researcher/engineers) part of human capital. **Research progress** seems to be the term with most orthogonality from fungible resources -- research provides discrete jumps and new paradigms while fungible resources tend to maximize current paradigms. One can therefore think of AI researchers as a proxy for "research progress potential."
It became immediately clear to me that there is an almost religiously set assumption that people have regarding how to model AI researchers. Specifically, people are polarized on what I call the **closed**/**open** axis²³, where **closed** is the direction of "only a few relevant people matter" and **open** is the direction of "the number of skilled workers matter."
For **closed** models:
1. Forecasting (at least on the research front) basically becomes an incentives game simulating people coming from the top labs (OpenAI, DeepMind, Anthropic) and how they are motivated to move around or cooperate. Would they be open to or allowed to work with China? What actions are they taking? For example, Meta is currently infusing massive amounts of money into the game to become an active player.
2. To give an example of how certain factors may become irrelevant for closed models:
1. Suppose for your purposes you only care about the top labs and top (\~50 per lab) AI researchers.
2. **Assuming that the top researchers preferentially go to the top companies,** then even a big change in immigration policy—say a 50% decrease on H1B caps from China—would probably result in only a minor change in the top Chinese talent a single lab would be able to access. These top researchers might be predominantly on O1 visas anyway.
3. In other words, **broad immigration policy would be less important in a closed model**, though of course there are always exceptions if the policy is drastic enough.
3. For an **extremely closed** model, I expect states to track the very top personnel (maybe O(5) to O(100) people at each leading company) and possibly monitoring/limiting/manipulating their moves. I have found that people using extremely closed models consistently favor the US in a "US vs China" race frame, because the top researchers are almost all in US-based institutions.
I think it is important to consider both open and closed models:
1. There are many mini-competitions in "the" AI race. For example, it is possible to believe both that "the race for AGI innovation" fits a very closed model but "whose LLMs have the most market share" fits a very open model²⁴.
2. Even in innovation, where people tend to lean towards closed or "a few geniuses" models, it is still very unclear whether ideas really beat scale in the long run (see e.g. The Bitter Lesson²⁵).
3. Even if the main players believed in an extremely closed model for their most important directives, they would **still** be trying to win quantities of researchers, wielding weapons such as immigration policy. This is essentially indistinguishable from them using a mix of the two models.
**At a Glance:** the main divide in how people model the importance of AI researchers can be placed on a closed/open axis, with the "closed" direction meaning that only a few researchers "really matter" and the "open" direction caring more about the quantity of talent. Different aspects of the "AI race" are different along this axis.
## **Open Models**
For **open** models, I think the two main factors to consider are **immigration** and an intangible that I call "**intellectual market share.**" If I spent more time, I would focus on items such as education/universities and domestic economies.
On immigration:
1. For a beautiful visualization of the worldwide movement of AI researchers in general, see the Global AI Talent Tracker²⁶. I think the two most important terms are:a. there's a heavy attractor to the US.b. Overall talent is "sticky" in the sense that most talent stays in the country of the grad school²⁷.
2. As US researchers tend to prefer to stay in the US, I think the main term to consider is **US immigration policy on Chinese nationals.** The main relevant visas for Chinese nationals in the US seem to be H1B (specialist) and O1 (extraordinary ability) visas.
1. The total number of H1Bs given is about \~65000 a year, of which roughly half would be in the relevant²⁸ "Professional, Scientific, and Technical Services" sector. Recently (not counting pandemic years) there have been 10000-20000 O1s issued a year.
2. One can use e.g.[ <u>https://h1bdata.info</u>](https://h1bdata.info/) to get a sense of counts of "elite" researchers on H1Bs. As a quick look: OpenAI has about 5000 employees and Anthropic has about 1000, of which reports 93 and 53 on H1Bs, respectively (a cursory glance shows that maybe around 80% of each are for "highly technical" staff). I was unable to find public information on how many O1s were at these labs.
3. If I spent more time, I would focus on EU, Russian, and Indian researchers.
Related to above, I have found that many people using open models are interested in a tight cluster of concepts that do not seem to have an agreed-upon pointer, so I will introduce the notion of **"Intellectual Market Share,"** loosely defined as "who do you think is leading AI research" or "if I want to e.g. do text-to-speech, what country's AI brands do I trust" or the pointed "I am a good but not top-notch researcher. Where do I see myself working?". I would estimate "Intellectual Market Share" through a few proxies:
1. **Publications** (not limited to academia; e.g. arxiv or conferences) seem the most accessible proxy variable for AI research, but private conversations with leading practitioners in industry have led me to strongly **downgrade** this as a good proxy, with the sense that much of the new progress is more time-lagged (from creation to conferences e.g. NeurIPS) or simply undisclosed / hidden.
2. **Competitions / Metrics:** comparative metrics such as the Livebench (for LLM capability), Scale SEAL (for LLM alignment/safety), or Artificial Analysis Arenas (for text conversion to other modes).²⁹
3. **Standardization** also seems important (akin to controlling terrain and can create competitive strength from entrenchment -- NVIDIA through CUDA (Compute Unified Device Architecture) is the best recent example of this).
1. I believe people overvalue DeepSeek as "winning" some contest of models (i.e. in the cost/efficacy tradeoff) but undervalue its "win" through the associated GitHub repos of related tools.
2. One ongoing battle of standardization of instruction set architecture, which has been historically dominated by x86 (mostly used in desktops, used by Intel and AMD) and ARM (mostly used in mobile). However, RISC-V (Reduced Instruction Set Computer - V) has been emerging as an open-standard, and China's incentives align with maximizing usage of RISC-V. This goal is easier if China gains "intellectual market share."
**At a Glance:** While it is not a market per se and hard to quantify, I think "intellectual market share" conceptually diversifies sources of "AI power" from our other inputs. For example, it would include the effects of academia and/or open source projects / standards. It gives a compass for attracting / ranking intellectual talent. The memes that "German cars are reliable" or "The US has the best schools" say more than the metrics that try to capture them do; they create very strong defaults in people's minds that affect their decisions significantly.
Predictions regarding human capital, disregarding how much it "matters for AI research":
1. I expect the US to be less of a desirable place for Chinese to come, but the baseline number of Chinese who want to come to the US will be high enough that overall this does not become a major factor.
1. I expect a moderate probability (40%) that there will be a significant specifically "anti-Chinese" immigration law³⁰ (such as a law preventing certain Chinese nationals or recently nationalized U.S. Citizens to exit the US or return to China) by 2030.
2. I expect the number of Indian engineers to increase in the US as a mild corollary.
2. On intellectual market share, my sense is that **the U.S. will still be dominant** but **China will gain ground by 2030**. As proxies:a. I predict a small but significant probability (15%) that China will beat the U.S. in Stanford HAI's (Human-Centered Artificial Intelligence) "most high-citation papers³¹" indices by 2030 (China already steadily leads³² both "most papers" and "most citations").b. I predict (65%) [<u>that RISC-V processors would reach 30% adoption</u>](https://www.metaculus.com/questions/39281/risc-v-30-market-share-by-2030/)³³ by 2030.
## **Data**
The primary relevance of data is **as a resource for AI training**. In order of estimated importance, Epoch AI³⁴ identifies energy, chips, data availability, and latency as four key possible bottlenecks for AI training scaling through 2030. We are now considering the possibility that good data for training will possibly "run out" around 2030, with some more aggressive estimates saying we are there already.
On Data:
1. \*\*Relevant types:\*\*
1. Right now, most of the relevant data is **text**, though especially with the rising demand of integration with other data modalities (text-to-video and text-to-speech), there should be rising demand for data in alternate formats, starting with images and audio.
2. It is possible that we end up in some world where the main breakthrough is to improve the quality of synthetically generated data³⁵ instead of needing to "find them in the wild."
3. Possibly the most important kind of upcoming data is **real-life environments**. While some of this will certainly be gathered as a part of robotics / military / automated vehicles research, the ability to generate synthetically generated RL (Reinforcement Learning) environments would be a very important battleground for industry and governments.
2. \*\*Data as state resources.\*\*
1. China's very centralized app ecosystem and surveillance mechanisms are preexisting forces to gather data, so AI increases their incentives to collect data even more. While the US nominally has more decentralized data ecosystems, e.g. the Snowden leaks show that at least some aspects of the state are interested in very large-scale data gathering, which the state might be eventually incentivized to more forcefully integrate with AI pursuits.
2. For data on data's availability on the state scale, see e.g. Tufts' Digital Planet research³⁶ on data pools.
3. **Data as industry resources.** Besides states, industries also keep "dark pools" of data (some might be at least partially regurgitated from running LLMs, so there is also a dimension of quality).
1. As data scarcity becomes important, states will be incentivized to tap into this data and/or protect it from other actors.
2. Data is easier to steal than hardware, so quick changes in data dominance is more likely compared to e.g. semiconductor dominance.
**At a Glance**: Data is an important resource that might spike in importance by 2030, probably due to alternate forms of data being demanded. In longer time-scales, I expect some game to emerge where "pools" of data are hoarded / shared / stolen as a strategic resource.
Predictions:
1. As a proxy for China's data strategy, I predict that China will start to leverage Africa as both a political dependent and data source for AI generation.
1. I expect (75%) cities in at least 12 countries in Africa to adapt to Chinese Smart City technologies by 2030³⁷ (9 as of 2020, though looser³⁸ counts already give 12).
2. I expect (70%) [<u>at least 50% of the smartphones in Africa to be Chinese-made</u>](https://www.metaculus.com/questions/39283/chinese-phones-50-in-africa-2030/)³⁹ by 2030. This is actually not that obvious since there's no obvious trend in the last 7 years -- Samsung + Apple still control cumulatively about 45% of the market -- so this is not a straight-line extrapolation but a bet on a spike.
2. There will be noticeable spikes of capital/activity on non-text data by 2030.
1. A proxy for video, audio, and music: we can take the McKinsey AI Report⁴⁰ for organizations, which as of 2025 sees (13%, 13%) for organizations report generating data for (video, audio/music) respectively⁴¹. I expect (80%) for those numbers to go up to at least (40%, 25%) by 2030.
2. A proxy for RL environments: by 2030, I expect (85%) at least 3B USD of confirmed investment / collaborations between leading LLM-based companies and "traditionally non-LLM" companies (video games, robotics, military) for the explicit goal⁴² of generating synthetic RL data.
# **Taiwan**
Many people wonder if China will attack Taiwan. After some tinkering to get a more resolvable question in the context of AI, I arrived at this question : **"**[**<u>By December 31, 2030, will there be a serious escalation about China committing to an action that, if completely successful, would control TSMC?</u>**](https://www.metaculus.com/questions/40242/prc-seeks-control-of-tsmcs-taiwan-ops/)**"** Examples of situations that I would count as positive resolution are:
1. land invasion (note that this still counts if the US chooses to destroy TSMC in a successful Chinese invasion),
2. large cyberattack that if successful would freeze/paralyze TSMC for an extended period of time (which could then be used as leverage to actually "take over" TSMC),
3. a direct surrender of power (or a coup) in a way that TSMC falls directly into Chinese governance,
4. The wording "serious escalation about" is to include cases such as the US preemptively⁴³ taking an aggressive action nominally to stop a committed Chinese action. Examples would be choosing to destroy TSMC themselves or a preemptive strike on Chinese military equipment.
The outline of my approach and updates are:
1. My basic model:
1. China wants "official" ownership of Taiwan for various reasons (TSMC, morale / legacy, control of surrounding waters).
2. The main non-militaristic-strength deterrence for China attacking Taiwan is potential economic damage (sanctions, blocking access to semiconductors, etc.).
2. Starting point (January 2024): I had my own (unpublished) estimate of 40-45%. A quick breakdown of the thinking process then:
1. As a starting point, a representative "consensus" of the time, a closely related Metaculus question⁴⁴ had a median of 26% of a *full-scale* invasion. Note this would be a **lower bound** for my question since I included other vectors.
2. There were various reasons I had an **increased** estimate from this baseline:
1. As an example of other vectors besides land invasion, there was (is!) an escalating invisible cyberwar at the time which had the intelligence community alarmed. Each individual alternative vector feels small. For example, semiconductor fabrication is a combination of software, hardware, human capital, and institutional knowledge that is very hard to steal (as compared to e.g. an LLM), so any such cyberattack that seeks *to control and not just to destroy* must be highly coordinated and sophisticated. On the other hand, "threaten to destroy" may be a good enough threat if it would not be seen as a bluff (this is made more effective by e.g. China's economy and chip sufficiency becoming more robust than projected). When combined, they feel like a **moderate increased estimate.**
2. I felt that Western forecasting/reports had a noticeable empathy deficit for the Chinese (both simulating individual Chinese people and the government's actions), and my mental models of Chinese leadership put much more importance on Taiwan as an objective aligning with other wants (for example, the morale gain would alleviate internal economic concerns, esp. fallout from COVID and 躺平 (tangping) social movements). This was not a huge increase because forecasters were already considering these things, but I felt they were under-considering the effects.iii. Evidence that China's military renovation (which was due to finish around 2029 in time for the National People's Congress) was going according to plan, which included investing significant observable resources in directly increasing capabilities of an invasion of Taiwan.
3. My updates since then:
1. As a baseline, the above Metaculus question's median has moved up 5% to 31% since 1/2024 as of now (6/25). It is unclear how much of the information they are using was already accounted for by my own delta with them from 2024, but it should by default make me move up somewhere between 0% and 5%.
2. Updates regarding the Trump administration:i. Its stated Taiwan stance seems much more uncertain than predecessors, though not enough to be surprising. This feels like a **small increase**, but is small precisely because Trump's words seem less reliable. On the other hand, Trump's unreliability also **decreases** the probability—it will make China wary of escalation unless it is sure of decisive victory. Overall, I want to disregard "cheap talk" signals as mostly noise.
1. Weapons sales seem hard to fake. A cursory glance shows that arms sales to Taiwan has been similar to recent presidents, including a sizable spike⁴⁵ in 2024. This **decreases** the probability.
2. The Trump Administration's increasing isolationist stances might lessen other actors' reliability in an alliance. For example, Japan + Netherlands + US can currently block China from making its own advanced fabs⁴⁶. This lessens the strength of this deterrence, which **increases** the probability.
3. The Russian invasion of Ukraine was less efficient/successful than median predictions, which **decreases**⁴⁷ the probability.
4. Semiconductors:
1. The failure of the Chinese chip self-sufficiency by 2025 is an indicator of less capacity/power, but would make Taiwan more strategically important. Intuitively, this **slightly increases** the probability of an invasion, but I also see counterarguments⁴⁸.
2. Diversification of chip manufacturing away from Taiwan decreases the "importance" of Taiwan to the US and thus **increases** the probability of an invasion. For example, maybe TSMC's fabs could be sufficiently (almost definitely only partially) duplicated in the US, which makes its loss during an invasion less costly to the US.
5. Other news, such as China building new ships optimized for Taiwan landing (which is a continuation of China making costly and specific steps for landing capacity)⁴⁹.
6. This aggregate puts it at 65%.
4. One more round of adjustment (overconfidence, forecasts of the type "will X happen" have the property that it is psychologically easier to find increasing signals than decreasing signals, etc.) gives a slight **decrease.**
5. I put my final estimate at 60%, which I think would be about 20-25% higher than Metaculus forecasters.
# **Other Predictions / Considerations**
I end with some scattered predictions and chains of thought, selecting for items that seemed especially interesting to myself and my discussion partners. These will be more speculative and less rigorous. Many of them would make for better prediction exercises after a few years. (while I thought it would be too much to outline my thinking for all of these, I am happy to discuss any of these individually)
On the EU:
1. I don't expect the EU to be a "major player" but I expect them to come up with a lot of ideas about safety regulations (with the goal of limiting the US and China).
2. I expect the EU's main strategy is to attempt to place themselves as a kingmaker, counting on ASML and Mistral to have a seat at the table (as well as EU's sizable investments⁵⁰ into the US, which it may be able to threaten to change given enough political will).
On Russia:
1. I expect China to be Russia's main ally and trading partner for the foreseeable future, especially with Russia's self-isolation from attacking Ukraine.
1. China's two main imports are chips (\~550 billion in 2024) and crude oil (\~360 billion in 2024). Russia desperately needs chips under sanctions, but consumes less than 1 billion in them yearly, so it is relatively cheap for China to give them chips.
2. I expect that China will absorb a sizable amount of Russian human capital into e.g. Huawei. See e.g. RUSI (Royal United Services Institute) research⁵¹. I expect in 20 years (if we are still around) and looking back, we would see a Russia → China "brain drain" that is comparable (though probably smaller) than the Russia → US "brain drain" during the cold war.
3. I might go as far to say a Chinese "vassalization" of Russia (or at least important parts of its economy / talent pool / etc.) in the medium-long term, though it might be hard to define.
2. It's not clear to me whether the US is trying hard to pressure Russia at the moment in the Russo-Ukraine war. If it does, I expect the US's main strategy (on top of sanctions) to be inducing Saudi Arabia to start an oil price war with Russia.
On the Iran / Israel conflict⁵²:
1. If this conflict is restarted or prolonged in a way that includes U.S. involvement, it probably increases the probability of a Taiwan invasion. The boring reason is that the U.S. would be more entangled. China would likely support Iran through various channels. If U.S. involvement is non-prolonged—for example, a quick Iran loss—it probably decreases the probability of a Taiwan invasion.
AI + Military:
1. I expect governments to continue to increase military R\&D for drone warfare significantly⁵³. One interesting direction is the 5% NATO spending goal⁵⁴ which NATO states might want to "use up" for AI research but "counting" it as defense.
2. **Drones** will become (already have) a buzzword associated with AI and relevant for getting world leaders / military / media scared, which increases capital flowing into AI.
Regulations:
1. I expect countries to all pretend to work on safety/regulation but the US and China will ignore them (while following them on paper and/or blaming the other side for not following them). I don't know how to turn this into a concrete prediction, however.
2. US Regulations: The recent passing of the Big Beautiful Bill (especially the reversal on not allowing states to regulate AI for 10 years) signals to me that the current administration is significantly worried about AI in some direction⁵⁵. My sense is the US will still broadly not-regulate the **capabilities** of the leading labs, though it might regulate **control** over the labs and their output.
3. Chinese Regulations: I expect very little Chinese regulations. Chinese citizens are by far the most pro- and comfortable with AI, and the regulation documents⁵⁶ I've read seem to focus mostly on not outputting "incorrect text" rather than limiting e.g. AGI capabilities.
4. I have recently updated upwards (though still bearish) that some politicians everywhere fear at least the **economic** consequences of advanced AI. My foggy prediction on this is that these items will turn into socio-economic policy instead of strong "shut it down" policies.
Job Market / Data Interactions:
1. I expect software engineering in the US to be at least 80% automated in 5 years by most metrics (lines of code, number of engineers needed in comparison to status quo, etc.).
2. I predict that with the squeezing of job markets and the demand for data, data labeling jobs will increase worldwide significantly, probably as a sink for less competitive engineers / data scientists, with a "higher tier" expert data labelers (possibly moonshining from academia / industry / etc.). Data labeling quality will be important so there will be some sort of "policing" on data quality.
3. I expect countries behind by most AI/tech rankings but with data pools (my specific guesses would be India and Indonesia) to evolve (at least temporarily) value / jobs from this gap, such as a caste of data-related jobs, or being a target by bigger players as sources of data in return for infrastructure.
Chinese Tech/AI Industry
1. I expect **Huawei**'s influence and capabilities to continue to increase. Their alignment with the Chinese state, ability to generate power / energy, ability to absorb talent (from Russia + elsewhere) and overall high effectiveness seems vastly underestimated by Western media (and they often seem like direct targets of US propaganda / aggressive actions, which is additional evidence that the US takes them seriously).
2. I expect China to dominate the robotics industry via some visible metrics (for example, some IFR (International Federation of Robotics) ⁵⁷ index on installations), both game-theoretically (due to their mineral and industry advantages) and due to their observed actions. I also separately predict the robotics industry market demand to increase way more than expected; this is caused by a union of thoughts from the Data section and also the demand for drones / automated warfare in the generally more chaotic/isolationist world.
US Tech/AI Industry
1. There will be a contest of power between governments and companies to manage data. Depending on who has more power:a. Governments may force companies to hand over or share data, either with the government itself or with other companies to consolidate power.b. Companies may force governments to subsidize or hand over control. For example, Microsoft could get better deals because everyone is using Windows.
2. The power imbalance here can be measured as a vague index of **how much of the government is actually run by AI**.
1. For example, we might get into a position where a significant amount of the actual textual output of internal memos are actually written by AI. The related discussion of "how much of code is written by AI" in either open source or a private company has already started.
2. As a proxy, we might be able to make predictions such as "by using , 70% of UK parliamentary speeches by 2030 will be labeled as mostly written by AI"
3. Less scrupulous companies might even put backdoors in the AI so they collaborate with each other and affect the goodwill of the government (For example, an LLM from BadAI, a malicious company, might slightly and probabilistically alter results / use deflective language of queries that are related to BadAI. I expect at least one major AI company to do this under the guise of "making AI make better decisions"). If done correctly this would be very hard to detect.
4. (long term) If much of the "internal querying" and decision-making of a government is done by AI (and possibly coordinated by being from the same company / affected by the same backdoor code) we might come to a time where **AIs will essentially be asked to make decisions on AI policy**...
On informational control / aggressive actions:
1. In general, I expect all agents (including powerful individuals or governments) to broadly take **more aggressive actions** in all spheres. The main reasoning:
1. Previously, the main deterrence for powerful individuals being caught while performing unpopular actions was the ability to call the evidence fake. With better multimedia AI, the cost of generating fake actions (and also calling real actions fake) is drastically lower. This induces relatively more fake actions being generated, which aids the plausible deniability of calling real actions fake.
2. This creates a positive feedback loop with the overall increase of both real and fake data and events, with a significant part created synthetically by AI. This already adds to the cost of truth verification. It will become more natural for information consumers to "give up" and spend less energy on truth verification, which makes it **even less costly** for actors to simply perform unpopular actions.
2. This gives a whole family of predictions in all spheres—use your imagination. For our geopolitical sphere, I expect more overt use of governmental brutality, more assassinations, more overt lying, etc. from all parties, including anti-state actors.
**Note: This draft was frozen for content as of July 25, 2025. Subsequent developments—particularly regarding mineral export controls and related policy changes—are not reflected here.**
***
# **Acknowledgments**
I completed this project with support from Open Philanthropy, which[ <u>provided assistance</u>](https://forum.effectivealtruism.org/posts/HPdWWetJbv4z8eJEe/open-phil-is-seeking-applications-from-grantees-impacted-by) after my original funder (the Future Fund) closed. I thank the grantmaking staff at both organizations for making this project possible, and Jacob Steinhardt for his recommendation to apply for the original grant.
While I appreciated the many conversations about this project, I thank Jean-Stanislas Delain, Jing Chan, Nuño Sempere, Timothy Abbott, Misha Yagudin, Sam Huang, Jacob Leibowitch, Alya Tabbott, and four anonymous practitioners from elite AI companies for deeper conversations, with special thanks to David Wagner for offering notably more time and feedback. I thank ChatGPT and DeepSeek for general vocabulary acquisition and idea gathering during the research phase. No LLMs were used for any part of the text or structural formation of this document during the writing phase.
The secondary sources that were not explicitly cited but very important to my mental models, in no particular order:
* Daniel Yergin - The Prize
* Matthieu Auzanneau - Oil, Power, and War
* Dmitri Alperovitch and Garrett Graff - World on the Brink (and by extension, almost the entire backlog of Geopolitics Decanted)
* Sarah Paine's lectures
* Peter Zeihan - The Accidental Superpower (and Zeihan's vlogs)
* John Mearsheimer's talks
* Erich Grunewald - Introduction to AI chip making in China
* Chris Miller - Chip War
* Kai-Fu Lee - AI Superpowers
* Elad Gil's blog posts
***
## **Footnotes**
¹ As an example, non-prediction market people might care about a question like "Is China more antagonistic against the US now versus 2020?" where "antagonistic" would be very intuitive to them. The gamified nature of prediction markets make very clear resolution criteria necessary and often adds clarity, but is frequently very distracting. In our example, the word "antagonist" above would be very constipated to define through observable data, so a prediction market would be forced to split the above question into 3-4 subquestions, with each subquestion involving separate hair-pulling about resolution conditions.
² This specific year is relevant for a few Taiwan-related factors, such as the timing of many reforms and modernizations (including military). I also feel (affected by AGI etc.) that I am extremely confused for any forecasts more than 5 years out.
³ It is very easy to be distracted by "is OpenAI really a part of the US" so I will just admit that I am making hidden assumptions like "the US remains powerful enough to have leverage over 'US companies' like OpenAI / Meta such that it makes sense to call them 'part of the US'", at least for the scope of this work. I also suspect that the states are going to mostly "think" in this way.
⁴ The "rest of the world," mostly the EU, have cumulative money on the order of magnitude of just one of these players, and other types of resources far lacking. Though of course items such as TSMC, ASML, EU policies, are certainly relevant. Related: also see "**Final Disclaimer.**"
⁵ "Monotonicity" means that the agents by default would want to maximize the resource. This excludes items such as regulations.
⁶ Using Holden Karnovsky's language ([<u>https://www.cold-takes.com/making-the-best-of-the-most-important-century/#the-competition-frame</u>](https://www.cold-takes.com/making-the-best-of-the-most-important-century/#the-competition-frame)), this is the "competition" frame versus the "caution" frame.
⁷ See e.g.[ <u>https://chipexplorer.eto.tech/</u>](https://chipexplorer.eto.tech/) to examine different segments and leading players.
⁸ See[ <u>https://semiengineering.com/breaking-down-the-ai-memory-wall/</u>](https://semiengineering.com/breaking-down-the-ai-memory-wall/) for a discussion of the "AI Memory Wall."
⁹ According to e.g.[ <u>https://www.csis.org/blogs/trustee-china-hand/legacy-chip-overcapacity-china-myth-and-reality</u>](https://www.csis.org/blogs/trustee-china-hand/legacy-chip-overcapacity-china-myth-and-reality), China currently holds e.g. 27 percent of manufacturing for 28-65 nm process node production.
¹⁰ The NITRD separates "Core AI" (their examples include machine vision, cybersecurity, neuromorphic computing architectures, chips optimized for neural nets, and... algorithms for computational linguistics) and "cross-cutting" AI (robots, data analysis/management ecosystem, human-machine interaction, cybersecurity search (as in CSP), and general neuromorphic computing). My sense is that there is going to be quite a bit of overlap. There will be about 1.5 billion put into "core AI".
¹¹ This is a very loose following of[ <u>https://cset.georgetown.edu/publication/chinese-public-ai-rd-spending-provisional-findings/</u>](https://cset.georgetown.edu/publication/chinese-public-ai-rd-spending-provisional-findings/), which uses Chinese MOF (财政部) statistics from 2018. The reference is broken, so I deduced that they probably used 2018年全国一般公共预算支出决算表 ([<u>http://yss.mof.gov.cn/2018czjs/201907/t20190718\_3303195.htm</u>](http://yss.mof.gov.cn/2018czjs/201907/t20190718_3303195.htm)) so can compare with e.g. 中央政府预决算 ([<u>http://www.mof.gov.cn/zyyjsgkpt/zyzfyjs/</u>](http://www.mof.gov.cn/zyyjsgkpt/zyzfyjs/)) for 2024.
¹²[ <u>https://en.wikipedia.org/wiki/China\_Integrated\_Circuit\_Industry\_Investment\_Fund</u>](https://en.wikipedia.org/wiki/China_Integrated_Circuit_Industry_Investment_Fund)
¹³ Stanford HAI AI 2025 report,[ <u>https://hai.stanford.edu/assets/files/hai\_ai\_index\_report\_2025.pdf</u>](https://hai.stanford.edu/assets/files/hai_ai_index_report_2025.pdf).
¹⁴[ <u>https://digital-strategy.ec.europa.eu/en/policies/european-ai-research</u>](https://digital-strategy.ec.europa.eu/en/policies/european-ai-research)
¹⁵ Energy consumption for LLMs can be split into training (one time per model) and inference (long-term based on adaptation). It is debated (e.g. [https://arxiv.org/pdf/2205.09646)](https://arxiv.org/pdf/2205.09646\)) which one costs more, but to not get caught up in the weeds, we can assume their proportions are roughly constant across models and our timelines (e.g. say "let's look at energy consumption through 2030 across all models"), then training costs is a good proxy for overall energy consumption.
¹⁶[ <u>https://epoch.ai/blog/can-ai-scaling-continue-through-2030</u>](https://epoch.ai/blog/can-ai-scaling-continue-through-2030)
¹⁷ There does seem to be an increasing trend in renewables (they account for 32% of global electricity supply last year). I feel with longer-term forecasting (say 20 years) this might be relevant, but given my short-term forecasting (5 years) I feel mostly ignoring renewables is fine. Someone with different models on power usage disruption might differ.
¹⁸[ <u>https://www.energyinst.org/statistical-review</u>](https://www.energyinst.org/statistical-review), 2024.
¹⁹ While I indeed just say that oil is less important for energy consumption, as one of the world's most important resources that drive geopolitics, it is still very relevant when considering the states' actions.
²⁰ See[ <u>https://www.iea.org/reports/global-critical-minerals-outlook-2025</u>](https://www.iea.org/reports/global-critical-minerals-outlook-2025).
²¹[ <u>https://www.energy.gov/technologycommercialization/articles/ai-tool-speeds-critical-mineral-hunt-boosting-us-supply</u>](https://www.energy.gov/technologycommercialization/articles/ai-tool-speeds-critical-mineral-hunt-boosting-us-supply)
²² Where "significant" is similar in spirit to the minerals prediction above. For energy, the actual methods might be something like power grid management or failure detection. See[ <u>https://www.iea.org/commentaries/why-ai-and-energy-are-the-new-power-couple</u>](https://www.iea.org/commentaries/why-ai-and-energy-are-the-new-power-couple) for ideas.
²³ I could have alternatively used "non-fungible" vs "fungible", but that would be confusing given my headings.
²⁴ Even the most ardent closed model believers I have talked to who would make bets like "China will make 0 major AI research innovations in the next 5 years" admitted that China can be more efficient at scaling, catching up after an innovation, etc. and is at least able to "win" metrics such as the AI consumer market.
²⁵[ <u>http://www.incompleteideas.net/IncIdeas/BitterLesson.html</u>](http://www.incompleteideas.net/IncIdeas/BitterLesson.html)
²⁶[ <u>https://archivemacropolo.org/interactive/digital-projects/the-global-ai-talent-tracker/</u>](https://archivemacropolo.org/interactive/digital-projects/the-global-ai-talent-tracker/)
²⁷ Also see e.g.[ <u>https://ncses.nsf.gov/pubs/nsf25325/table/1</u>](https://ncses.nsf.gov/pubs/nsf25325/table/1) for corroboration. However, I feel this needs some additional notes: there were giant shifts of STEM hiring (e.g. huge tech layoffs circa 2022-2023) and many of these studies are based on earlier data. Furthermore, from personal conversations and anecdotal observations, I am fairly confident that the US is becoming a less desirable place for Chinese workers to stay because of standard of living and increasing tensions. I suspect the effect size of this to be small but nontrivial on top AI researchers.
²⁸[ <u>https://ssti.org/blog/useful-stats-look-h-1b-visa-program-industry-employer-and-state</u>](https://ssti.org/blog/useful-stats-look-h-1b-visa-program-industry-employer-and-state)
²⁹[ <u>https://artificialanalysis.ai/text-to-video/arena?tab=leaderboard</u>](https://artificialanalysis.ai/text-to-video/arena?tab=leaderboard) and two others for text-to-speech and text-to-text.
³⁰ Could be included in export controls, which might be moving in the direction of forbidding movement. For example there was a rule in the 10/7/2022 US export controls restricting the ability of US citizens or green card holders to not work at PRC-located semiconductor facilities.
³¹[ <u>https://hai-production.s3.amazonaws.com/files/hai\_ai\_index\_report\_2025.pdf</u>](https://hai-production.s3.amazonaws.com/files/hai_ai_index_report_2025.pdf) I looked at the reports for 2019 through 2025 (though 2020 seems to have been skipped). China already comfortably leads both the "number of papers" and "citations" indices, but not yet the "number of high-citation papers" index of publications selected for having citations above a high threshold, which the U.S. is currently winning.
³² Some of every part of academia is susceptible to insular cliques of citation, but there's no a priori reason to believe that the "citation inflation" here differs significantly from the status quo and I suspect there is fairly strong bias against Chinese papers by outsiders, or even insiders who do not work with research. Talking to AI practitioners on the research side, I anecdotally (n \< 10) find that the quality of Chinese papers seem comparable. Of course, all of this is under the caveat that the most relevant results are probably currently unpublished.
³³ As defined by e.g. Omdia[ <u>https://omdia.tech.informa.com/om121430/risc-v-global-market--2024-analysis</u>](https://omdia.tech.informa.com/om121430/risc-v-global-market--2024-analysis). I am counting on China's push for RISC-V adoption for my higher estimates.
³⁴ [https://epoch.ai/blog/can-ai-scaling-continue-through-2030](https://epoch.ai/blog/can-ai-scaling-continue-through-2030)
³⁵ As a mental experiment, we can imagine a human writer as someone that turns data into "new" text, so there is probably some level of "quality" after which we would accept the output of AI text as "interesting" enough to be consumed by other AIs. For a speculative guess at the "critical point", we might imagine AIs interacting with each other in a simulated RL environment and accumulating "life experiences" that are then turned into text; one can imagine the richness of such output to be "sufficiently novel."
³⁶[ <u>https://digitalplanet.tufts.edu/the-emerging-geography-of-ai/</u>](https://digitalplanet.tufts.edu/the-emerging-geography-of-ai/)
³⁷ [https://www.csis.org/analysis/chinas-smart-cities-africa-should-united-states-be-concerned](https://www.csis.org/analysis/chinas-smart-cities-africa-should-united-states-be-concerned)
³⁸ [https://www.csis.org/analysis/watching-huaweis-safe-cities](https://www.csis.org/analysis/watching-huaweis-safe-cities)
³⁹ According to e.g. [https://gs.statcounter.com/vendor-market-share/mobile/africa](https://gs.statcounter.com/vendor-market-share/mobile/africa)
⁴⁰ [https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai#/](https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai#/)
⁴¹ They also have "other data" at 2% which I expect to go up, but as not all sectors would use e.g. RL environments seriously, I am less confident about that number. While I do assume total RL environment data to drastically increase, it might be focused on a few large players.
⁴² Thus, I would not count someone working with the military for improved drone production, but I would count it if part of the output would be to reuse the drone-captured data to train AIs.
⁴³ We can imagine something like "US intelligence receiving intel that China is about to strike in the next week" as a positive resolution. The veracity of the claim will be impossible to verify, but I think this is an extremely important part of the probability space and accounts for a big part of my heightened estimations.
⁴⁴ [https://www.metaculus.com/questions/11480/](https://www.metaculus.com/questions/11480/)
⁴⁵ See CFR, [https://www.cfr.org/article/us-military-support-taiwan-five-charts.](https://www.cfr.org/article/us-military-support-taiwan-five-charts.)
⁴⁶ World on the Brink.
⁴⁷ Important note: compared to the counterfactual of predicting "Russia doesn't attack Ukraine" in 2022, I believe the Russian invasion increases the probability of a Taiwan attack, but here I am tracking additional updates already having known about the Russian invasion. For example, one of the biggest terms is "land warfare seems more difficult than anticipated."
⁴⁸ Note that such a term would flip for a forecaster who e.g. believes the US can 100% reliably blow up TSMC and 100% reliably blockade semiconductors from China w/ an alliance.
⁴⁹[ <u>https://www.navalnews.com/naval-news/2025/01/china-suddenly-building-fleet-of-special-barges-suitable-for-taiwan-landings/</u>](https://www.navalnews.com/naval-news/2025/01/china-suddenly-building-fleet-of-special-barges-suitable-for-taiwan-landings/), etc.
⁵⁰ The EU is responsible for 45% of investments into the US; see e.g.[ <u>https://www.visualcapitalist.com/who-invests-in-the-u-s-the-most/</u>](https://www.visualcapitalist.com/who-invests-in-the-u-s-the-most/) for 2023. This does not account for sectors, so I looked at[ <u>https://www.bea.gov/sites/default/files/2024-07/dici0724.pdf</u>](https://www.bea.gov/sites/default/files/2024-07/dici0724.pdf) which had a more detailed breakdown (though of cumulative numbers). My conclusion is the heuristic is good (though interestingly there are some cancellations; for example, the EU is underrepresented in "computer equipment" (around 25%) but then overrepresented in "information" (around 66%).
⁵¹[ <u>https://www.rusi.org/explore-our-research/publications/commentary/hi-tech-high-risk-russo-chinese-cooperation-emerging-technologies</u>](https://www.rusi.org/explore-our-research/publications/commentary/hi-tech-high-risk-russo-chinese-cooperation-emerging-technologies).
⁵² To clarify: I do not mean the Iran-Israel 2025 War that (literally literally) just started, but the prolonged conflict (with other conflicts such as Palestine serving as proxies), similar to the stance of the Institute of Study of War[ <u>https://www.understandingwar.org/</u>](https://www.understandingwar.org/).
⁵³ c.f. The US DOD's massive increase in contracts in 2023 (from \~350M to 4.6B USD!
⁵⁴ [https://www.nato.int/cps/en/natohq/topics\_49198.htm](https://www.nato.int/cps/en/natohq/topics_49198.htm)
⁵⁵ For transparency: I would have put at least 70% against this happening, so I would have been incorrect.
⁵⁶ See e.g. 生成式人工智能服务管理暂行办法[ <u>https://www.gov.cn/zhengce/zhengceku/202307/content\_6891752.htm</u>](https://www.gov.cn/zhengce/zhengceku/202307/content_6891752.htm) for a representative example.
⁵⁷ [https://ifr.org/worldrobotics/](https://ifr.org/worldrobotics/)