[Weekly] US-Japan Gap Market Map #003 (2026.7.20-7.26) — Capital shifts from 'Above AI' to 'Below AI'. Buyers are picking up plumbing, prep work, and cleaning jobs

Series Purpose

This letter is a series edited for individuals who want to buy small digital assets / want to buy them someday by observing the market that has just emerged in the US before it arrives in Japan. The intended readers are those who want to eventually become buyers of 'businesses within reach of individuals,' such as SaaS with monthly sales of several hundred thousand yen, dormant apps that have stopped updating, and membership businesses. This is not a technical news source for creators, but a map for buyers.

Each issue answers the following three questions: (1) How many years will the gap in this market last? (2) Who will buy the businesses created there in the future, and why? (3) If an individual wants to become a buyer, what should they be looking at now? The material is one week's worth of the 'Time Machine Management Daily' (14 reports total: 2 systems x 7 days), which tracks startups, funding, and product trends in the English-speaking world daily. All figures are limited to public and reported values stated in the original reports; no new estimates are added here. This is not a recommendation to buy or sell specific assets, nor is it investment advice. Please use it strictly as a 'workbook for reading the market as a buyer'.

This week's conclusion (3 lines)

First, this week the center of gravity for capital shifted entirely from 'Above AI' (models/apps) to 'Below AI' (the control layer of payments, identity, auditing, and inference). US VC investment in the first half of the year reached a record high of $412.7B, with over 81% being mega-rounds of $100M or more, indicating an 'oligopolistic market.' Second, the most flashy flow of this massive capital was into Physical AI (robots, chips, nuclear reactors), which is 'out of reach' for individuals, but serves as a preview of what might fall into the second-hand market in five years. Third, our assessment this week is that what individuals can realistically pick up are the three mundane jobs created around this massive trend: 'plumbing' (infrastructure for the agent economy), 'prep work' (structuring messy data), and 'cleaning' (tidying up the chaos scattered by AI).

Feature 1 (Meta News) — Capital shifts from 'Above AI' to 'Below AI'. The control layer became the biggest battlefield this week

1. What happened in the US

The most significant fact this week is not about individual products, but the shift in 'which layer the money flowed into.' US VC investment in the first half of 2026 reached a record high of **$412.7B, of which over 81% were large rounds of $100M or more (7/26 report 2). The market has clearly entered a 'mega-oligopoly.' And the destination for that massive capital was not flashy generative AI apps, but the 'lower layers' for running AI. Inference cloud Together AI raised $800M in Series C for a valuation of $8.3B, with annual bookings reaching **$1.15B**, and its valuation increased 2.5x in 18 months (7/23 report 2). AI materials discovery firm CuspAI raised $450M in Series B with a valuation of $2.6B, with investors including Bezos Expeditions, Nvidia, AMD Ventures, and even the UK government (7/22 report 2). On the security side, non-human identity (NHI) firm Oak exited stealth with a total of $60M (7/21), AI app/agent defense firm Neo raised $100M (led by a16z, 7/22), and endpoint defense firm Glow exited stealth with $180M and a valuation of $1.2B (7/23-24). Agent payment firm Natural raised $30M to compete with Stripe (7/21-22), and inference optimization firm Infinity raised $15M** (7/21).

2. Why did this happen?

The reason is the cost structure. The performance of the models themselves has become commoditized, and there is no longer a difference in 'which model to use.' The difference lies in the part that safely connects AI agents to actual business operations—who approves payments, what permissions they operate under, how the audit trail is kept, and how cheaply inference can be run. Once integrated, this is hard to rip out and has low churn. That is why investors started betting on the 'lower' plumbing rather than the 'upper' apps. The CEO of Hugging Face stated this week that 'companies have stopped renting AI,' which points in the same direction; it means the experimental phase is over and we have entered the stage of integrating it as permanent infrastructure. The language of capital has changed from the era of flashy chatbots to the era of mundane instruments and plumbing.

3. Asymmetry on the Japanese side

In Japan, this 'control layer' is structurally empty. Domestic security products (IAM/EDR) are still designed on the premise of 'PCs used by humans,' and the concepts of permission management, activity logs, and audit trail presentation when an AI agent acts on behalf of a human are weak. Moreover, it is common for Japanese corporate information systems departments to have only a few dozen people, making it physically impossible to manually inventory the ever-increasing 'Shadow AI' (AI tools introduced by departments without permission). On the payment side, Japan has almost no credit, retry, or fraud detection mechanisms that support a world where 'machines pay'—where agents automatically buy goods or APIs. In terms of systems, J-SOX (Internal Control Reporting System), the revised Personal Information Protection Act, and Financial Services Agency guidelines impose accountability on companies for 'who did what with which authority,' and the more AI agents are introduced into operations, the more demand for tools that present those audit trails will automatically arise. There is a gap here where 'legal clauses generate revenue.'

4. Implications from a buyer's perspective

This is the main point. An individual cannot buy a $100M defense SaaS, but they can pick up a 'scaled-down version' of this trend. Specifically, there are two ways to act. First, look at small security SaaS or operational log tools already running domestically (log management, permission inventory, sender domain authentication, etc.) with the eye of 'can this be horizontally expanded to audit trail presentation in the AI agent era?' If it is a lightweight version for Japanese SMEs in a field where the US is seeing $100M valuations, it falls within an individual's reach as an asset with monthly sales of several hundred thousand yen. Second, re-translate the operational know-how you already possess—such as payment failure recovery or email deliverability management—into the context of 'plumbing for the agent economy.' Payment retries are necessary not only for 'human payments' but also for 'machine payments,' and email authentication is continuous with the layer that detects 'naturally written AI-generated phishing.' You are buying existing small assets, but the source of the valuation has shifted to 'how long will this continue to be needed in the agent era?'

5. Observation indicators to watch

First, adoption of agent payment/identity standards. Follow whether Visa/Mastercard or platformers officially adopt 'payment/identity verification protocols for AI agents.' The moment the standard is solidified, small jobs like Japanese localization and SME implementation will be mass-produced. Second, speed of mandatory Shadow AI control in Japan. Whether the Ministry of Economy, Trade and Industry or the Financial Services Agency will explicitly define 'preservation of AI usage audit trails' as a de facto obligation. If it is mandated, the unit price of compatible tools will rise. Third, movements toward OEM supply to SIers/MSSPs (Managed Security Service Providers). Since it is heavy for an individual to sell to companies alone, the resale value of small assets in this area depends on 'whether it is in a form that can be OEM-supplied to major SIers.' You cannot buy massive funding rounds like Together AI or CuspAI, but it is worth looking for 'Japanese plumbing parts' that fall at their feet while they are still cheap.

Feature 2 (Regulatory Driver + Pickaxes) — Laws have put price tags on 'codes for regulated industries' and 'agent auditing'

1. What happened in the US

On July 25, a coding agent for regulated industries 8090 raised a Series A of $135M. Led by Salesforce Ventures, the CEO is Chamath Palihapitiya (7/25 report2). Their selling point is using AI to modernize legacy systems in industries like finance, insurance, and public sector—where mistakes lead to legal liability—while meeting audit requirements. In the same week, capital concentrated on agent 'defense' and 'audit trails.' AegisAI raised $36M to launch a defense that analyzes AI-generated spear phishing and BEC (Business Email Compromise) for every incoming email (7/24 report3). In addition to the aforementioned Oak (NHI, $60M), Neo ($100M), and Glow ($1.2B), Empirical Security raised **$25M** (total $37M, 7/22). France's ANSSI (National Cybersecurity Agency) announced that starting in 2027, it will block certification for products that do not support Post-Quantum Cryptography (PQC) (7/22 report3). In other words, this is a move where the law forces demand, stating that 'if you don't update your encryption, you can't sell.'

② Why it happened

In a world where AI agents perform tasks on behalf of humans, being able to prove 'who approved it, with what authority, and what they did' after the fact becomes just as important as the execution itself. While not flashy, this is an area with the qualities buyers love most: demand created by the force of regulation, solid contracts, and low churn. It is symbolic that 8090 specialized in regulated industries; the takeaway this week was that while 'writing code with AI' is a commodity, 'writing code in a way that withstands audits' commands a high price. Defense, auditing, and cryptographic migration are all on the 'pickaxe' side, where work increases as AI becomes more widespread.

③ Asymmetry on the Japanese side

Japan has a double void in this area. One is the massive amount of legacy systems sleeping in regulated industries (banks, insurance, local governments). COBOL and old core systems are still active, and because modifications involve Financial Services Agency inspections and internal control requirements, they are too burdensome for ordinary SIers and are left untouched. The other is the absence of a layer that automates those modifications and operations 'in a way that withstands audits.' J-SOX, the Electronic Book Preservation Act, and the invoice system have effectively mandated electronic evidence of vouchers and operations, but there are almost no domestic tools that leave and present this as 'AI agent action logs.' While domestic IT departments are exhausted by manual inventory, there is no player claiming to be the receptacle that takes over that burden with AI. Regulations like ANSSI's 'forced cryptographic migration' are still in the discussion phase in Japan, with no implementation.

④ Implications from a buyer's perspective

What individuals can pick up are these 'plain parts of regulatory compliance.' First, look at small compliance/audit SaaS already in use domestically (access log management, voucher storage, sender domain authentication, etc.) to see if they can be rebranded with the new sales pitch of 'evidence presentation in the AI agent era.' Second, look for small membership businesses or outsourcing targets in niche regulated industry tasks (document conversion for financial institutions, document processing for insurance agencies, etc.) that can be taken over by AI. While large capital like 8090 goes after the main targets in regulated industries, small automations for 'one industry, one document' that fall through the cracks are areas individuals can secure early and cheaply. The point is whether the revenue is based on 'customer sentiment' or 'legal statutes.' Statute-based models generate renewal demand with every institutional revision, have low churn, and are assets that are easy for buyers to price.

⑤ Observation indicators to watch

First, strengthening of evidence requirements for domestic regulations. Will the Financial Services Agency, METI, and the Personal Information Protection Commission explicitly state 'log retention and accountability when using AI' in their guidelines? Second, domestic schedule for Post-Quantum Cryptography (PQC) migration. If moves like ANSSI's to 'exclude non-compliant products' spread to Japanese government procurement, a new market for cryptographic migration outsourcing and auditing will emerge. Third, statistical data on AI-generated phishing damage. As the amount of BEC and phishing damage published by the National Police Agency and IPA grows, budgets will be allocated to Japanese versions of 'AI-detecting' defenses like AegisAI. None of these are suggestions of 'what to buy,' but rather indicators for observation—when the statutes move, the voids will increase in value.

Special Feature ③ (Japan's Void) — 'Prep work' before AI acts. Structuring messy data will be the main battlefield in 2026

① What happened in the US

On July 24, XCures, which structures messy medical records, raised a Series B of $46M. It is a company that organizes 'messy data'—such as doctors' handwritten notes, medical charts in disparate formats, and test results—into a form that AI can use (7/24 report2). On the same day, a16z's Jennifer Li stated that 'structuring messy data is the main battlefield of 2026.' No matter how smart AI models become, they cannot provide value if the input data is messy—so capital is heading toward 'prep work before AI acts.' This trend was observed simultaneously in multiple industries. Mercura (YC W25), which handles wholesale and manufacturing orders with AI, structures non-standard orders that come in via fax, phone, and Excel (7/20-22). Accounting firm Bluebook (YC W25) announced 'AI-native accounting,' where AI handles everything from vouchers to journal entry decisions (7/25 report3). These are all examples where value was placed not on 'flashy AI,' but on the gritty maintenance work that precedes it.

② Why it happened

Over the past two years, the AI narrative has focused on 'smart models.' However, as companies actually try to extract value from AI, it has become clear that the biggest bottleneck is not the intelligence of the model, but the messiness of internal data. When medical charts are fragmented by vendor, orders come in via fax, and vouchers are scattered across paper and Excel, no matter how excellent the agent is, it cannot act. That is why a company that 'just organizes' gets $46M. This has properties that are convenient for buyers. Once data structuring takes hold, it enters the root of operations, and the more it is used, the more data assets—such as 'company-specific name-matching dictionaries and conversion rules'—are accumulated. These accumulated data assets become barriers to entry that competitors entering later cannot easily imitate.

③ Asymmetry on the Japanese side

Japan is one of the countries with the deepest mountains of this 'messy data' in the world. In medicine, electronic medical records are fragmented by vendors like Fujitsu and NEC, and data formats are not standardized between facilities. B2B transactions for SMEs are still dominated by fax, phone, and Excel ledgers, and EDI (Electronic Data Interchange) has only spread to standardized transactions of large companies. SME ordering, which centers on high-mix low-volume and custom orders, is a void for structuring. The '2024 problem' in logistics has led to an absolute shortage of manpower, making it impossible to leave this inefficiency unaddressed. In accounting, freee and Money Forward have only reached the stage of 'making input easier,' while the actual judgment of journal entries is still handled by humans, and the tax accountant industry that handles those judgments is aging. While the invoice system and the Electronic Book Preservation Act have effectively mandated the digitization of vouchers, the layer that organizes that electronic data into a 'structure that AI can use' is completely empty.

④ Implications from a buyer's perspective

This may be the most realistic hunting ground for individual buyers this week. The reason is that the target can be carved out into 'one industry, one document.' While a massive platform that organizes data for all industries is a battlefield for large capital, small SaaS or outsourcing businesses that 'structure specific documents for a specific industry' can be operated by individuals. Specifically, there are two ways: (a) buy small outsourcing businesses (data entry, document digitization, bookkeeping, etc.) that are up for sale due to succession difficulties, under the premise that 'profit margins can be increased by structuring with AI,' or (b) in a field where you have industry knowledge (e.g., ordering or application documents for a specific industry), use your knowledge of the 'habits of how it is messy' as a weapon to accumulate assets in the form of conversion rules. You are buying a plain administrative business, but the source of pricing has shifted to 'how much prep work can be automated with AI' and 'how difficult the accumulated conversion data is to imitate.'

⑤ Observation indicators to watch

First, 業種特化の構造化データの取引価格. On business succession M&A platforms (Tranbi, Batonz, etc.), check how 'seemingly plain outsourcing businesses' like data entry or bookkeeping are being bought and sold in terms of monthly sales multiples. If this starts to trend upward under the premise that 'it can be lightened with AI,' it is a sign that the void has begun to close. Second, 電子カルテ・EDIの標準化政策. When the Ministry of Health, Labour and Welfare's standardization of electronic medical record information and policies for the spread of EDI for SMEs move, the need for structuring will suddenly become apparent. Third, インボイス・電帳法対応の実務負担の声. The more the dissatisfaction that 'vouchers have been digitized, but are ultimately being organized manually' is visualized on the front lines of SMEs, the more the value of small assets that fill that gap will rise.

Mini Feature — The Financialization of Collectibles. Despite being IP originating from Japan, the exchanges are being built overseas.

As this week's 'early observation of a new category birth,' I want to record the movement to buy and sell trading cards and collectibles like financial products. In the reports for July 20 and July 26, Misprint (YC W25) was featured for allowing Pokemon cards to be traded on an order book 'like Robinhood,' and the existing player Alt (a collectibles exchange) has reached a valuation of over $800M. The mechanism is simple: it brings price visualization and real-time order book trading to trading cards, which until now had their prices determined by over-the-counter purchases and the shopkeeper's eye.

There is a strong asymmetry here. Both Pokemon cards and Yu-Gi-Oh! are IP originating from Japan, and the domestic trading card market is said to be worth approximately 300 billion yen. Nevertheless, their secondary distribution is still centered on storefronts and flea market apps, and there is no 'exchange where market prices are visible on an order book' in Japan. Although the appraisal infrastructure (grading) is being put in place as PSA's Japanese base begins operations, the financial trading layer that sits on top of it remains a blank space. From a buyer's perspective, this is a reenactment of a pattern that Japan, a content powerhouse, has repeated: 'assets originating in Japan are being financialized overseas first.' While individuals cannot suddenly create an exchange, there is room to enter on a small scale in the periphery—the behind-the-scenes work such as appraisal, authenticity verification, inventory management, and the supply of price data. Including the fact that the permit design of the Secondhand Articles Dealer Act is the key to entry, this is a category worth observing to see 'who would build it if a Japanese version were created, and which parts would be sold first.'

Watchlist (10 items)

1. Autonomous Dental Clinic Operation OS (US Denta / YC S26). Japan has about 67,000 dental clinics, more than convenience stores, and the shortage of hygienists, the mandatory electronic billing for receipts, and the high volume of phone appointments make it favorable for AI automation. Existing systems have fragmented receipt computers and web booking, with no layer to run operations across the board. To a buyer, this looks like the seed of a 'small tool that lightens the entire administrative burden of a single clinic.'

2. Operational Foundation for Clinics Led by Nurses with Specific Training (US Corner Health / $32.5M). While NP-led models themselves are prohibited in the US under Article 17 of the Medical Practitioners' Act, Japan has a 'Nurse with Specific Training' system (38 acts in 21 categories), and there is no software to support procedure manual management, logs of comprehensive instructions, or remote verification. The cap on overtime work for doctors is a tailwind. Entry difficulty is somewhat high; for now, it is an observation target.

3. Inheritance Tech (End-to-End Estate Settlement) (US Elayne / YC S24, Empathy total over $100M). Although there are about 1.5 million inheritances per year with a transfer value of 50 trillion yen annually, collecting family registers and inheritance division agreements are completely analog. The digitization of the statutory inheritance information list and the wide-area issuance of family registers are tailwinds, and the shortage of judicial scriveners and tax accountants drives demand. The boundary of the professional services law is the key.

4. AI Email Security (Defense against Japanese BEC/Spoofing) (US AegisAI / $36M). Domestic products focus on sanitization and preventing misdirected emails, leaving a blank space for 'detecting natural Japanese phishing generated by AI through context.' There is demand in finance, manufacturing, and local governments, and it is adjacent to the persistent themes of this series (email deliverability and sender domain authentication).

5. Real-time Japanese Voice AI Infrastructure (French Gradium / $100M, Nvidia participation). Domestic speech synthesis focuses on broadcasting and narration, and lacks the interruption tolerance and sub-hundred-millisecond response required for conversational agents. There is demand in call centers, local government counters, and nursing records. Since this is a capital-intensive area, it is an observation target for the time being.

6. Real-time Voice Translation API for Developers (US Pinch / YC W25). In the era of 40 million visitors to Japan, even if there are dedicated devices like Pocketalk, there are almost no APIs that can be embedded into apps, POS, and reservation systems. There is room to pick up small embedded demand.

7. AI-Native Accounting (Automating up to Journal Entry Judgment) (US Bluebook / YC W25). freee and Money Forward only go as far as input assistance, leaving journal entry judgment to humans. The aging of tax accountants and the mandatory digitization of vouchers due to the invoice system and electronic bookkeeping law are tailwinds. Since professional services law is a barrier, buying a bookkeeping agency business is a realistic form.

8. Automation of Non-Standard Ordering for Wholesalers and Manufacturers (US Mercura / YC W25). Japanese B2B is still dominated by fax, phone, and Excel ledgers, and EDI is only for standardized transactions of large companies. Small and medium-sized enterprises' high-mix, low-volume, and custom orders are a blank space, and the '2024 Logistics Problem' is increasing the urgency. If you narrow it down to a specific industry, it can be operated even by an individual.

9. AI Role-Play Corporate Training (US Synthesia / Valuation $2.1B). While the Japanese corporate training market is about 500 billion yen, group training and one-way e-learning videos are mainstream, and role-playing is operated in a person-dependent manner. It is highly compatible with honorifics, complaint handling, and customer service, and can be established as a small, industry-specific membership business.

10. Apps within ChatGPT (Japanese Local Use) (OpenAI Apps SDK / 800 million users per week). Japanese consumer apps are centered on proprietary apps and the App Store, and there are no ChatGPT-native examples. For uses such as transit, gourmet reservations, administrative procedures, household accounts, and learning, a design that turns into the 'side called by AI' is also continuous with AEO (AI Search Optimization).

Numbers of the Week

• $412.7B / Over 81%: US VC investment amount in the first half of 2026 (all-time high) and the ratio accounted for by mega-rounds of $100M or more. The market has entered 'oligopoly'.

• $1.15B: Annual bookings for inference cloud Together AI. Valuation 2.5x in 18 months ($8.3B). A symbol that capital has moved 'below AI' = the control layer.

• $135M: Series A for 8090, a coding agent for regulated industries (Chamath CEO). The price tag attached to 'code that can withstand an audit'.

• $46M: Series B for XCures, which structures messy medical records. Proof that 'prep work before AI moves' has become the main battlefield.

• 800 million people per week: Distribution scale of ChatGPT Apps SDK. The new entrance for apps is shifting to the 'side called by AI'.

• $55.8B: Broad Physical AI investment in 2026 (YTD). It has surpassed the 2021 peak of $14.1B with half a year still remaining.

• 15 million people: Number of registered users for the post-algorithm SNS 'Yope' (the majority use it over 5 days a week). Disengaging from the attention economy is becoming a consumer trend.

This week's 'Things You Can't Buy'

Huge news with a 5-star entry barrier is not for individual shopping. Even so, it is worth keeping in the corner of your eye as a trailer for 'what kind of used assets will hit the market in 5 years.' This week, that was most evident in the shift of AI bottlenecks from 'chips → power → physical (Atoms).' Atoms, the industrial Physical AI company that Travis Kalanick has been developing in stealth for 8 years,Atoms raised **$1.7B** (with Uber also investing, 7/22-26), targeting everything from mining and construction to heavy transport and food production. The trigger is investment in VLA (Vision-Language-Action) models, which serve as the 'brains' of robots. On the power side, nuclear fission startup Valar Atomics entered funding negotiations at a valuation of $6 billion (approx. 900 billion yen) (7/20-21), and in chips, Etched reached a valuation of $10.3B (approx. 1.5 trillion yen) for its Transformer-dedicated inference chip (7/24-26). Data infrastructure company Databricks has a valuation of **$188 billion** (approx. 28 trillion yen), and synthetic biology company Colossal is in discussions at **$20-30 billion**. These are all worlds where the numbers are off by about four digits, but this capital shift from 'models to chips, power, and physical' will push small software and membership businesses that fell through the cracks into the secondary market once the consolidation cycle is complete. You can't buy the huge news, but the used assets that fall downstream are the hunting grounds for individuals.

Three questions for buyers this week

① How many more years (or months) will this market gap last? The three gaps this week—plumbing for the agent economy, audit trails for regulatory compliance, and structuring messy data—are all 'just-started' markets in the US that have begun to see $30M–$180M valuations. There is a time lag in systems and business customs for landing in Japan, and that difference is your grace period for preparation. Try placing your watch-list items to see which of these three they connect to.

② Who will buy the business created there in the future, and why? The answer pattern this week was clear. Small SaaS for agent defense is bought by large SIers/MSSPs to 'fill holes in their own managed services,' and data structuring contract work is bought by industry platforms to 'take over the maintenance of their own data.' Whose roadmap hole is your candidate filling? Assets where you can identify the exit (buyer) first are the ones you can acquire cheaply and sell reliably.

③ If an individual is to become a buyer, what should they be looking at now? There were three filters this week. (a) Is the revenue based on 'legal statutes' rather than 'customer mood' (audit obligations, crypto migration, invoices, etc.)? (b) Are you standing on the side that 'increases invoices' due to the spread of AI, rather than the side that gets 'swallowed' by the evolution of AI models (plumbing, auditing, and prep work are the latter)? (c) Does it possess data assets that accumulate the more they are used (conversion rules, name-matching dictionaries, detection logs)? Small assets that satisfy these three are the types that tend to rise in value quietly, especially in an era of oligopolistic markets.

Disclaimer

The figures in this article are public values and reported values stated in source reports; no new estimates have been added here. This does not recommend the buying, selling, or investing in specific assets, nor is it investment advice. Market outlooks are the author's editorial judgment, and actual decision-making should be based on your own research.

Next Issue Preview

#004 (2026.7.27-8.2). From the 14 items next week, I will again edit 'markets before they come to Japan' from a buyer's perspective.

Link to BiP Series

Deep dives into individual cases (implementation notes on the perspective of actually 'buying and running' things, such as payment recovery/retainer flow/email deliverability) are handled continuously in the BiP Magazine series. Let's meet with both a buyer's eye and a creator's hand. Please check it out.

I run a community that studies how to take profitable overseas businesses, arrange them for the Japanese market, build them without failure, and grow them. From candidate selection → numerical verification → Japanese differences → manual verification → MVP → pricing → expansion. Click here for LINE registration: https://lin.ee/8Yyw7wk