AI for Science Detailed Analysis (Part 2): Capital and Landscape

Introduction: From Technology Landscape to Capital Landscape

The previous article covered the technology landscape of AI for Science: which directions have already emerged, which are still stuck in papers and prototypes, and which are more like long-term visions. After looking at the technology landscape, the next question to ask is about capital: Is anyone in this field willing to keep pouring money in? Where is the money going, and where is it avoiding?

Technology reviews focus on system capabilities; the capital perspective concerns industrial judgment. A demo that looks impressive might not attract bets. An unremarkable tool-layer company might be raising significant funds. Capital isn’t necessarily smarter than technology, but it reveals one thing: which paths the market believes can become a business, and which paths are still just research hype.

This article shifts the axis to look at AI for Science (hereafter using the same abbreviation AI4S from the previous article). My judgment is simple: money is indeed coming in, the door hasn’t closed on this track, but opportunities are not evenly distributed. The top layer requires top-tier teams and hundreds of millions of dollars to start—out of reach for the average person. The lower layers, such as tools, evaluation, and vertical infrastructure, still have entry points. Distinguishing between these two layers is more useful than broadly asking “Is AI4S worth doing?”

Chapter 1: Macro Background of Capital: A Highly Concentrated New Financing Regime

To see how capital enters AI4S, we first need to look at the venture capital landscape from 2025 to 2026. AI4S isn’t raising funds in isolation; it coincides with a period of intense AI fundraising, which is also highly unbalanced.

1.1 AI Absorbs Most of the Money, and It’s Extremely Concentrated

First, look at the totals. In 2025, roughly half of global venture capital funding went to AI, with the U.S. market approaching 60%. By Q1 2026, global VC funding was about $300 billion, of which AI companies took about $242 billion, nearly 80%. At this ratio, AI isn’t just a hot direction; it’s redistributing the entire VC market’s money.

The issue is that this money isn’t flowing evenly to all AI companies. In 2025, about 40% of U.S. venture capital funds were concentrated in the top ten companies. OpenAI received approximately $122 billion in committed capital in March 2026, pushing its valuation to around $852 billion. Anthropic raised about $65 billion in May, reaching a valuation of about $965 billion, surpassing OpenAI to become one of the highest-valued frontier AI model startups. xAI raised about $20 billion in Q1 2026, merging with SpaceX in February, with a combined valuation of roughly $12.5 trillion. Waymo raised about $16 billion in February, primarily from its parent company Alphabet.

So, current AI funding resembles a dumbbell. A few top-tier companies swallow massive amounts of capital; at the other end are numerous small, early-stage startups. The middle segment is thinning out. Specifically, early-stage small rounds in the millions of dollars range are becoming rarer. There appears to be more money, but it’s not for everyone; it’s flowing to the extremes.

Looking at deal counts makes it clearer. In 2025, AI took about half of VC money but only about 30% of deal count. This means the average deal size for AI companies is much larger. Meanwhile, small early-stage rounds under $5 million dropped to their lowest share in nearly a decade. For the average entrepreneur, this means the market is hot, but getting a decent first check is actually harder than in previous years.

1.2 Who’s Writing the Checks Has Changed

Where the money goes has changed, and so has who’s writing the checks.

Traditional VC funds can no longer support the multi-hundred-billion or trillion-dollar rounds of companies like OpenAI and Anthropic. Two new types of backers have stepped in. One is sovereign wealth funds, like Abu Dhabi’s MGX, Qatar’s QIA, and Saudi Arabia’s PIF. Saudi Arabia’s PIF alone invested about $36.2 billion in AI in 2025. The other type is strategic investments from tech giants. In OpenAI’s $122 billion committed capital round, Amazon, NVIDIA, and SoftBank were strategic anchor investors, with Microsoft continuing to participate.

The goals of this money also differ from traditional venture capital. When NVIDIA invests in a company, financial returns are a consideration, but it’s also securing future computational power customers. When sovereign funds invest, they aim to tie their countries to the most cutting-edge AI technology and talent. So these mega-funding rounds are more like collaborative arrangements between large corporations and state capital, a different game from the risk capital familiar to startups.

This is crucial for understanding AI4S. At the most capital-intensive model layer, who survives is no longer mainly decided by traditional VCs like a16z and Sequoia, but by tech giants and state capital that already have money, compute, and channels. Traditional VCs need to find other positions, and AI4S offers one such new position.

1.3 What This Macro Environment Means for AI4S

Placing this backdrop into AI4S reveals that the macro environment and company realities aren’t perfectly aligned.

From a macro perspective, AI4S is riding the AI funding boom: plenty of hot money, high attention, and top-tier institutions are looking. But at the company level, core AI4S players are mostly very young, many just founded, still at seed or Series A. Their individual funding rounds are completely incomparable to the hundreds-of-billion-dollar rounds of model companies.

This discrepancy has a direct consequence: the bar for early-stage funding has risen. When the entire market’s money chases a few star companies, an ordinary early-stage AI4S company finding its first decent check is harder than a few years ago. Those that can raise significant money either have extremely strong team resumes or hit upon a very hard capital logic. That logic is the “data depletion” topic we’ll cover next.

Chapter 2: Data Depletion: Why Top Capital is Willing to Invest in AI4S

The previous article mentioned a background fact: high-quality internet text is diminishing. This chapter places that background into the capital context. Only by understanding this can you grasp why institutions like a16z and NVIDIA would invest in a few newly formed scientific labs that don’t even have a product yet.

2.1 Internet Text is Running Out, Experiments Are the Next Data Goldmine

The starting point is simple: training cutting-edge large models requires high-quality data, and the available human-generated text on the internet is being consumed. Some research estimates that usable human-written text could be exhausted as early as 2026. When “feeding text” hits this ceiling, the next high-quality data source becomes the industry’s most pressing question.

Many investors and entrepreneurs are pointing to real-world scientific experiments. Periodic Labs puts it directly: data is finite; state-of-the-art AI models have used up most of the available data in recent years, and every experiment they run, success or failure, generates several gigabytes of new data not found on the internet. Co-founder Liam Fedus (a co-author of ChatGPT) put it more bluntly: the next step is to bring experiments into the AI loop, letting models interact with the real world.

Why experimental data specifically, rather than scraping more websites? Because experimental data offers two things internet text cannot. First, it is causal: a real-world experiment records “doing this yields that result,” a causal chain precisely what plain text lacks. Second, it has built-in right and wrong: whether a material is qualified, whether a reaction succeeds—the physical world provides an undeniable judgment. This kind of definitive feedback signal is the rarest fuel for training next-generation reasoning models. In short, the internet teaches models to talk; experiments teach models to be correct.

In investment terms, this means: scientific experiments are no longer just the work of scientists; they can also become a data source for large models. Whoever can stably produce physical, chemical, or biological data not available on the internet may hold the raw material for the next wave of model training.

2.2 How Data Depletion Changes Capital’s View

This change alters how capital views AI4S.

In the past, “using AI to help science” was more of an academic and philanthropic topic, about papers, discoveries, and long-term value for humanity. Now, “data depletion” ties it directly to the next step in large model development. Investing in a company that can build its own lab and produce exclusive experimental data is no longer just supporting scientific research; it’s hoarding a strategic resource unavailable elsewhere. The slow business of science thus gains a story that can appeal to AGI investors.

Taking a step further, science labs start to resemble data companies. Papers, materials, and molecules are surface products; the underlying proprietary experimental data is the harder-to-replicate asset. Especially as “verifiable rewards” become more important, a laboratory that can continuously provide real-world feedback is very attractive to model companies and investors. This is why several science labs that haven’t yet productized can raise hundreds of millions of dollars.

But this story can’t be overblown. “Data depletion, experiments are goldmines” has a real logic, but also fundraising rhetoric. Founders and investors are motivated to make it sound more attractive because this narrative can support a multi-hundred-million dollar valuation for a company without products. Calmly, at least two questions remain unanswered. First, how much of this experimental data is truly scarce, high-signal data that can make models stronger, versus just repetitive experiments and low-information noise? Second, whether scientific experimental data can train stronger general models is still more a belief than a proven fact. Read this article with these two question marks.

This explains the change mentioned earlier. The model layer is already occupied by tech giants and state capital. Traditional VCs need to find new positions. Science companies that can generate exclusive experimental data provide such an entry point.

2.3 Two Derived Logics: Selling Shovels and Digging Moat

Around “data depletion,” two more specific investor preferences have formed.

First, investors often prefer infrastructure over directly building an “AI Scientist.” Directly building an AI Scientist is high risk, burns cash fast, and has a long cycle. Providing the foundation, data, evaluation, and tools for scientific agents has more certain demand. Not every gold prospector strikes it rich, but the shovel sellers are more likely to get paid. The same logic applies in AI4S.

NVIDIA is the most typical infrastructure player. It invested in Periodic, also in Lila, and participated in the NSF’s open science large model project. No matter which science company eventually succeeds, if they need massive compute, they’ll have to go through NVIDIA. This also reminds us not to simply equate “NVIDIA invested in company X” as a pure quality endorsement. It certainly indicates the company is worth attention, but also involves ecosystem positioning and locking in compute clients.

Second, there’s a preference for finding truly defensible vertical domains. As the previous article described, when “making an AI application” becomes increasingly easy, what’s truly valuable is what others can’t replicate: proprietary data, switching costs in customer workflows, and know-how accumulated over years of disciplinary work. Capital in AI4S is looking for exactly these positions that are not easily flattened by the underlying model.

Chapter 3: Representative Fundraising and Player Landscape

Now that we’ve covered why capital is willing to invest, let’s see where the money is actually going. The following companies and projects broadly represent the main directions of AI4S fundraising from 2025 to 2026.

3.1 The God-Building Layer: Building Own Labs, Creating “AI Scientists”

The most eye-catching and capital-intensive category involves building your own lab, aiming for a closed loop between AI and automated experiments.

The representative company is Periodic Labs. Founders Liam Fedus (ex-OpenAI, co-author of ChatGPT) and Ekin Dogus Cubuk (ex-DeepMind, led GNoME materials model) started the company in 2025, focusing on using automated physical labs to run experiments independently, producing new materials and new data. It received about $300 million in a seed round led by a16z, with NVIDIA’s investment arm, Bezos, Schmidt, etc., valuing it at around $1.3 billion. Reports indicate it’s later discussing a new funding round that could push the valuation to around $7 billion, though this is not finalized. A seed round of $300 million for a very early-stage company shows how much capital is willing to pay for this narrative.

Another is Lila Sciences. It was incubated from the Flagship Pioneering biotech investment system in 2023, publicly aiming to create “a superintelligence for science.” It debuted in March 2025 with about $200 million, then a Series A raised approximately $235 million, followed by another $115 million, totaling about $350 million for the A round, with NVIDIA’s investment arm also participating. At this point, its total funding exceeds $500 million, with a valuation around $1.2 to $1.3 billion.

If we narrow it down to AI materials discovery, related startups have collectively raised over $1.3 billion in the past two years.

These fundings bet on the same thing: whoever can smoothly run the “AI design + automated experiment” loop might be able to produce new materials in batches and accumulate experimental data others can’t get. But this path is heavy and risky. Labs must be built, equipment purchased, scientists hired—the burn rate rivals training large models, and results won’t be visible for several years. This is important when discussing who can enter the playing field later.

3.2 IP and Milestone Model: No Lab, Selling Designs

Not every company needs to build its own lab. There’s a lighter approach: only do AI design, selling the designed molecules, proteins, or enzymes as intellectual property to pharmaceutical companies, collecting milestone payments.

The most typical example is the Profluent and Eli Lilly partnership. In April 2026, AI protein design company Profluent entered a strategic collaboration with Eli Lilly. Profluent uses AI to design site-specific recombinases for gene medicine, with Eli Lilly receiving exclusive rights. Profluent is eligible for up to $2.25 billion in milestone payments.

The advantage of this model is clear. Profluent doesn’t need to build factories, run clinical trials, or navigate regulatory hurdles. It provides only the core AI design capability. The heavy, slow, and expensive development is left to a big pharma company like Lilly with money, channels, and regulatory experience. The AI company designs, the pharma company provides funding and execution capability—this is a pragmatic path for AI4S in life sciences.

It also has another benefit: transferring part of the timing risk to the pharma company. Clinical trials take years, regulatory hurdles must be overcome—things that take time and burn money are handled by Lilly. Profluent only needs to continuously produce good designs upfront to collect milestone payments. For teams with professional capabilities but unable to bear heavy assets and long cycles, this path is worth serious consideration.

3.3 Evaluation and Tool Layer: Shovel Sellers Also Raise Big Money

Now look at tools and evaluation. This layer is often underestimated—many think “scoring models” or “making tools” is just supporting work, not as sexy as directly building an AI Scientist. But funding data shows these companies can also raise substantial amounts.

The best example is LMArena. It spun out from UC Berkeley’s crowdsourced scoring project Chatbot Arena, specializing in large model evaluation and leaderboards. It raised about $100 million in a seed round in May 2025, valuing it at about $600 million, led by a16z and the University of California investment office. By January 2026, it raised about $150 million in Series A, with valuation rising to about $1.7 billion. An evaluation platform accumulating about $250 million shows that evaluation itself can become a big business.

Another example is Axiom Math, founded by Carina Hong from Stanford, focusing on mathematics and formal verification. It raised about $64 million in a seed round in October 2025, and about $200 million in Series A in March 2026, valuing it about $1.6 billion. Although the name includes “Math,” it publicly emphasizes “proving that AI-generated code is safe.” Pure math research is only part of it; the commercial selling point is closer to software verification.

3.4 Catalytic and Government Capital: Underwriting Where VCs Won’t Go

Another important but often overlooked type of money in AI4S comes from government and philanthropic institutions. They don’t necessarily seek short-term returns, instead underwriting infrastructure and early-stage research.

Google.org has an “AI for Science” funding program, about $30 million, focusing on health and life sciences. The U.S. Department of Energy launched the “Genesis Mission” via executive order in November 2025, aiming to mobilize national labs to accelerate scientific discovery with AI, followed by a tender of about $293 million. The National Science Foundation and NVIDIA jointly launched the OMAI project (Open Multimodal AI Scientific Infrastructure). NSF contributed about $75 million, NVIDIA about $77 million, totaling about $152 million, handed to the Allen Institute for AI (Ai2) to build fully open large models for science.

Beyond government, the Schmidt Sciences Foundation (former Schmidt Futures) of Eric and Wendy Schmidt has long funded AI and science cross-disciplinary research, like the AI2050 Scholars program. NSF’s own “National AI Research Institutes” program has invested over $500 million cumulatively. The common characteristic of this money is that it doesn’t require short-term commercial returns, making it more suitable for supporting basic research and public infrastructure. For academics wanting to enter AI4S, this is often more suitable than direct venture capital.

Simultaneously, top VCs are moving towards hard tech. a16z raised a roughly $15 billion fund in January 2026, the largest in venture capital history, with a dedicated slice of about $1.176 billion for defense, aerospace, manufacturing, energy—their “American Dynamism” direction. Capital is moving from light software towards harder technology domains, and AI4S is in this wave.

Putting the companies and projects from this chapter together looks roughly like this:

(Note: The original article includes a table summarizing key players, funding, rounds, etc. The table structure should be preserved in translation, but the markdown for the table is not explicitly provided in the user message. Based on the description, it likely lists: Periodic Labs [300M seed], Lila Sciences [500M+ total, multi-round], Profluent [milestones up to 2.25B], LMArena [250M seed + A], Axiom Math [264M seed + A], Google.org [30M grant], DoE Genesis [293M tender], NSF+NVIDIA OMAI [152M], Schmidt Sciences [???], NSF National AI Institutes [500M+]. I will construct a reasonable markdown table based on the text.)

Company/Project	Focus	Key Funding ($)	Stage	Notable Backers
Periodic Labs	Self-driving lab, materials/data	~300M (seed)	Seed	a16z, NVIDIA, Bezos, Schmidt
Lila Sciences	AI scientist for biotech	>500M (total)	Early	Flagship, NVIDIA
Profluent	AI protein design (IP licensing)	Up to 2.25B milestone	Partnership	Eli Lilly
LMArena	LLM evaluation/leaderboard	~250M (Seed+A)	Early A	a16z, UC
Axiom Math	Math formal verification	~264M (Seed+A)	Early A	(not specified)
Google.org AI4S	Grant funding	~30M	Grant	Google.org
DoE Genesis Mission	National lab AI for science	~293M (tender)	Govt	US Govt
NSF+NVIDIA OMAI	Open science AI models	~152M	Govt-Industry	NSF, NVIDIA, Ai2
Schmidt Sciences	AI for science research	Ongoing	Philanthropy	Schmidt family
NSF National AI Institutes	Multi-institute AI research	>500M	Govt	NSF

Several names appear repeatedly among investors: a16z, NVIDIA, a few sovereign funds, a few government agencies. AI4S is currently far from a mass entrepreneurship track; the direction is being defined by a few top institutions. The downside is high barriers, making it difficult for ordinary teams to squeeze in. The upside is the landscape isn’t set yet; early entrants still have a chance to define a niche direction.

Chapter 4: Is the Track Still Early? Looking at Founding Years

Next, we need to address a more practical question: Is AI4S still early? Is it a new direction just getting capital attention, or is it already saturated with top players?

To judge this, don’t just listen to founders’ stories. Look at two hard indicators: when the core players were founded, and what round they are currently raising.

4.1 Founding Year is a Strong Signal

Whether a direction is new or old can often be seen from the “age” of its leading companies.

If the most prominent companies in a direction were founded in the last year or two and are still at seed or Series A, it indicates the territory has just been discovered, and positions aren’t filled. Conversely, if star companies have been around for many years, are generally at Series D or E, with valuations in the tens or hundreds of billions, it usually means the landscape is relatively stable, making it very difficult for newcomers to enter.

So when I look at a new direction, I don’t look at slogans first. I look at whether the top players are “young companies with early rounds” or “old companies with late rounds.”

This indicator is hard to fake. Companies can package their technology, media can chase trends, founders can pitch visions, but founding years and funding stages are relatively hard data. If a direction is already a red ocean, capital would have already fed the top companies to late rounds; it’s unlikely that core players are still collectively at seed and Series A.

4.2 Core Players Are All Very Young

Using this ruler to measure AI4S, the picture is clear.

The core players mentioned in Chapter 3 are mostly very young. Periodic Labs was founded in 2025, still at seed. Lila Sciences founded 2023, still early. LMArena spun out from a university project in 2023, just at Series A. Axiom Math founded 2025, just at Series A. They all appeared between 2023 and 2025.

The only slightly older one is Profluent, founded 2022, but it follows a lighter model collaborating with pharma and hasn’t reached a large late round. Including it doesn’t change the conclusion: among the core AI4S players, none has weathered five or eight years and grown into a behemoth. This timeline aligns with the judgment in the previous article: the technology landscape of AI4S truly took shape around 2024, and capital concentrated entry afterwards.

This indicates AI4S is still in an early stage. The best positions aren’t fully occupied yet, and capital is still competing for early-stage projects. Entering AI4S now is not too late in terms of timing; many sub-directions haven’t formed fixed patterns.

4.3 Control Group: What a Mature Track Looks Like

It’s not enough to say AI4S is young; it’s best to compare with a significantly more mature AI vertical.

Abridge is a good control. It focuses on AI medical note-taking in healthcare, automatically transcribing doctor-patient conversations. Abridge was founded in 2018 and had been around for seven years by 2025. It completed a $250 million Series D in February 2025, valued at $2.75 billion. Four months later, it completed a $300 million Series E, valuation rising to $5.3 billion. Its contracted annualized revenue for Q1 2025 was about $117 million.

Comparing Abridge and Periodic, the difference is clear. Abridge is seven years old, at Series E, has a $5.3 billion valuation, and over $100 million in contracted revenue. Periodic is very young, seed stage, with product and commercialization still on the way. AI4S is still years away from Abridge’s level of maturity.

Abridge also reminds us that an AI vertical track from early to mature typically takes several years. It was founded in 2018, with valuation truly taking off around 2025—six or seven years in between. Applying this timeline to AI4S, the conclusion is clear: entering now isn’t too late, but don’t expect results within a year or two. This is a multi-year marathon.

4.4 It Aligns with the 2026 Capital Trend

One more external signal is telling: top capital is moving towards hard tech.

As mentioned, top VCs like a16z are pouring big money into hard tech, with American Dynamism as a typical banner. Capital is moving from light software and SaaS (relatively crowded) towards defense, manufacturing, energy, and research infrastructure—harder areas requiring deeper expertise. AI4S is right in this wave.

For those wanting to enter, the key here is timing. AI4S hasn’t reached its end; it’s still early. The advantage of being early is that many positions are still open, and many sub-directions haven’t been defined yet. The downside is higher risk; commercialization paths may not be immediately clear.

However, “early” doesn’t mean “easy.” The previous article gave a self-check for moats, equally applicable in AI4S: Do you have proprietary data others can’t get? Are you embedded in the research workflow so that users find it hard to switch? Do you have know-how accumulated over years of disciplinary work? The more “yes” answers, the stronger your position. Next, we’ll discuss that although AI4S is early, not every layer is suitable for ordinary teams.

Chapter 5: Layer-by-Layer Judgment: God-Building Layer vs Infrastructure Layer

AI4S being early is good news. But early doesn’t mean everyone can do it. The barriers between layers differ significantly and must be examined separately.

Chapter 3 divided by funding type: God-building, IP & Milestones, Evaluation & Tools, Government Capital. Here we switch to a more practical division: Can an ordinary team actually enter? Simply put, the God-building layer is very hard to enter; the Evaluation & Tools layer is more realistic; the Profluent-style IP & Milestone model sits in between; government and philanthropic funds provide early-stage underwriting.

5.1 God-Building Layer: New, but Out of Reach

The first layer is the God-building layer—companies like Periodic and Lila that build their own labs and aim to create AI Scientists.

This layer is indeed new, indeed hot, and backed by top-tier capital. But for the vast majority, it’s unrealistic. Its entry ticket is a hall-of-fame caliber team, e.g., former core researchers from OpenAI and DeepMind, plus hundreds of millions of dollars in starting capital. Without this resume and funding, even the first round is hard to close. It can be viewed as a benchmark, but it’s not the path most teams should choose.

Moreover, despite their flashy fundraising, this layer carries the highest risk. They are simultaneously betting on three things: that AI can genuinely produce valuable scientific results autonomously, that automated labs can run stably, and that the experimental data generated can in turn train stronger models. If any one of these fails, hundreds of millions of dollars can burn down ugly. For most people, this layer is better suited for observation than for rushing into.

5.2 Infrastructure Layer: A More Realistic Entry Point

The second layer is the infrastructure layer—providing the foundation, evaluation, tools, and data for scientific agents. For ordinary teams, this layer is more realistic.

It’s more realistic because it doesn’t require a celebrity team and hundreds of millions of dollars from the start. A small team, by cutting narrowly enough, can find a position. Two directions here are particularly worth watching. One is Agent infrastructure, including evaluation, tools, orchestration. LMArena’s ~$1.7 billion valuation already shows that evaluation alone can become a big business. The other is deep vertical domains, like physics, materials, astronomy—directions requiring specialized judgment. AI engineering skills are important, but often can’t compensate for subject matter expertise.

To be more specific: For evaluation, you could create a truly professional exam for scientific agents in a particular discipline, like Gravity-Bench or ReplicationBench mentioned in the previous article, where people who understand the field set the standards and judge correctness. For deep vertical tools, you could connect specific professional data, a set of simulation code, or a type of experimental equipment into the Agent workflow, allowing it to actually do useful work in a narrow domain. Neither requires hundreds of millions to start; they require specialized judgment others don’t have.

This also ties back to the main thread of the previous article. The voice of the God-building layer lies with giants and top-tier teams. But at the infrastructure layer of evaluation, tools, and deep verticals, subject matter expertise becomes more important. For people with disciplinary backgrounds, this is the wall others can’t easily climb over.

5.3 Business Models: From Heavy to Light

From heavy to light, AI4S has roughly several different business models.

Heaviest: building your own lab, like Periodic and Lila, burning money to generate data. Lighter: IP & Milestones, like Profluent, doing AI design yourself, passing subsequent development to pharma. Lighter still: pure software tools, like LMArena and Axiom, mainly relying on products, evaluation systems, and data. Even lighter: open-source first, then commercialize, like Ai2 taking government and NVIDIA money to build open science models, building the ecosystem first, then finding commercial outlets.

The lighter the end, the more suitable for teams with limited capital but strong professional judgment. In other words, first look at what you have, then decide which layer to stand on.

Chapter 6: Risk Side and Exits

We’ve covered money and opportunities, but AI4S is not an easy path. It has several hard constraints that can’t be ignored.

6.1 Slow: The Return Cycle of Science is Very Long

First: slowness.

The business of science doesn’t operate on the same rhythm as customer service agents or coding assistants in enterprise AI. A customer service agent might see revenue in a few months, with clear feedback. But an AI-designed drug or an AI-discovered material takes years from design to validation to market, passing through experiments, clinical trials, and regulatory hurdles. Chapter 4 mentioned Abridge, which deals with relatively fast medical records and took seven years to reach a $5.3 billion valuation. Real hard science will only be slower.

This is a real test for capital. Venture capital funds have a lifespan; they need exits and returns. But science doesn’t progress according to fund cycles. Slowness is the most fundamental risk in AI4S, rooted in the pace of the business itself.

6.2 Exits Primarily via Acquisition, IPOs Slow

Second: exits.

AI4S companies are more likely to exit through acquisition than to grow up and go public independently. A small company with a unique molecular design capability has a more realistic destiny of being acquired by a big pharma like Eli Lilly or by a tech giant. Public market exits are slow, and samples are few. This means that money invested in AI4S largely waits for industrial giants to buy it out—a narrower path than internet software.

This isn’t entirely bad. The ceiling for acquisitions is usually lower than IPOs, but the speed might be faster, and the certainty higher. For a science team that has built something real but doesn’t plan to become a giant itself, being bought out by a big pharma or tech company for a good price is a decent outcome. The prerequisite is that what you have is hard enough that giants prefer to buy rather than build from scratch.

6.3 The “Science” Evaluation Sub-Sector is Still Early for Pure VC

The third problem lies behind the statement “evaluation is important.”

Evaluation as a broad category has been validated as profitable; LMArena’s ~$1.7 billion valuation is proof. But LMArena evaluates general large models with a wide audience and clearer commercialization path. Niche down to “science evaluation”—like creating exams and acting as judge for scientific agents in a specific discipline—the audience shrinks dramatically. Currently, this segment is mainly underwritten by government and philanthropic funds like the DoE, NSF, and Google.org. Science evaluation is real as an academic and infrastructure endeavor, but as a pure commercial VC business, it’s still early.

This connects with the judgment in the previous article. Evaluation is critical and suitable for people with disciplinary backgrounds. But be clear: it’s currently more like a field cultivated first by public interest and academia; the commercialization inflection point hasn’t fully arrived. The next article will discuss more specific approaches: not necessarily starting with a pure VC entrepreneurial route; you can start with research, evaluation, building reputation, then let opportunities come to you.

6.4 One More Thing to Watch For: Giants May Enter at Any Time

The final risk is that giants may enter the field themselves. Frontier model companies and big pharma have incentives to bring AI4S capabilities in-house. OpenAI and Google can build their own science teams; pharma companies like Lilly can develop their own AI design teams. Anthropic has already started pushing Claude into research workflows; Claude Science, an AI workbench for scientists, is a signal. Once these companies get serious, thin-shell companies built on “I know a little AI” will be the first to be squeezed out.

So it comes back to the moat. To survive under the gaze of giants, you need either proprietary data others can’t get, or you’re already embedded in customer workflows, or you have barriers built from years of disciplinary accumulation. Before entering AI4S, the thing to figure out most is not “Can I make a demo?” The real question is: Is my wall hard enough?

Summary: Money is Coming In, But Choose Your Position Right

To wrap up this article. Capital is indeed entering AI4S, and seriously. Core players are generally young, funding stages are early, indicating the landscape isn’t fully set. a16z, NVIDIA, sovereign wealth funds, the DoE, and the NSF are all investing in different ways.

But this is not a direction anyone can just charge into. The top God-building layer requires top teams and hundreds of millions of dollars; ordinary teams can’t touch it. More realistic positions lie below: evaluation, tools, data, deep vertical workflows. These places value subject matter expertise more than capital being the primary barrier.

So, for someone with a disciplinary background but without massive capital, the question shouldn’t be “Is AI4S worth doing?” The more specific question is: What identity should I enter with? Which layer should I cut into? Which narrow but hard problem should I solve first?

The next article covers this. Where are the remaining open spots? How can people from physics and other disciplines turn their strengths into irreplicable moats? Stay tuned.

Related Articles

• AI for Science Detailed Analysis (Part 1): Paradigms and Landscape

Selected Other Articles by the Author

• Grandfather Integral Calculus
• What is “Emergence”? A History of Emergence Research
• A Brief History of the Internet Bubble
• AI Circle Influencer List (List Expanding)
• My Personal AI System: Philosophical Foundations
• How the NFT Narrative Collapsed
• What is Dissipative Structure Theory? Does it Relate to AI?
• What is Embodied Intelligence? Its Relationship with AI
• In-depth Analysis: SpaceX Future Outlook
• How Vibe Coding Crashed My System: Summary and Takeaways
• One Article to Explain the U.S. Immigration System
• One Article to Explain the U.S. Healthcare System
• Detailed Analysis of the U.S. Chinese Old Money Families
• What Resources Did Jewish and Chinese People Get Respectively in the U.S.? Detailed Analysis
• One Article to Understand the Full U.S. Education Ecosystem
• What is Cybernetics? Is Cybernetics the Predecessor of AI?
• Introduction to Grandfather Integral
• Pope Leo XIV on Artificial Intelligence (Essential Edition)
• Vibe Reading: A Systematic Method for Reading in the AI Era
• Complete Guide to the U.S. Tax System

Main Data Sources

Organized by topic for easy cross-referencing. Capital figures are time-sensitive; please refer to the primary sources listed below for formal citations.

Macro Capital (Chapter 1)

• OpenAI ~122 billion committed capital, valuation ~852 billion: OpenAI announcement.
• Anthropic ~65 billion, valuation ~965 billion, surpassing OpenAI: AP, 2026-05-28.
• xAI merged into SpaceX, combined valuation ~$12.5 trillion: CNBC, 2026-02-03.
• Waymo ~$16 billion funding: CNBC, 2026-02-02.
• Top ten companies took ~40% of VC funds: SaaStr citing PitchBook.
• Q1 2026 record, AI accounts for ~80% of global: Crunchbase News.
• Saudi PIF, Abu Dhabi MGX AI deployments: CNBC, 2025-10-15.

Data Depletion and God-Building Layer (Chapters 2, 3)

• High-quality text to be exhausted around 2026: Nature.
• Periodic Labs ~$300 million seed round, founding team, “data creation” narrative: TechCrunch, a16z.
• Lila Sciences total funding >500 million, valuation ~1.2-1.3 billion: FierceBiotech, Axios.
• AI materials companies raised >$1.3 billion in two years: MIT Technology Review.

**