11 Years, $11B—What Comes Next? | A Conversation with Airwallex’s 吴恺: In the AI Era, the Next Stop Is $100B

Summary

• Airwallex’s latest funding round valued the company at $11B, underpinned by $1.3B in ARR growing 90% and industry gross margins of roughly 70%. CRO 吴恺 says the multiple is “not high—not high at all,” and “actually relatively low” after factoring in growth. The funding is primarily aimed at accelerating the recruitment of strong talent and sourcing high-quality data; acquisitions are focused on 2 things: people and data.
• The episode’s sharpest investment call is that AI could reshape vertical financial SaaS. 10 formerly separate categories—including expense management, accounting, revenue recognition and billing—could consolidate into a single agent interface over the next 5 years: “5 finance employees using 10 different software products could become 3 finance employees using 1 agent system.” Data fragmentation across point SaaS products makes automation fragmented and limits AI: “Its AI is relatively dumb.”
• Airwallex is betting on a financial super-intelligence entry point. 吴恺 rejects the view that super-intelligence will exist only at the general-purpose layer, arguing that proprietary, non-public data in each major industry creates a separate moat. Potential entrants include Airwallex itself and Stripe, which is “definitely building it”—its product lineup has expanded sharply over the past 3-4 years for precisely this reason. Revolut, with tens of millions of consumer users, is another potential player; “there aren’t many others.”
• Companies taking large models overseas—Kimi, MiniMax and 智谱—are creating a new business in real-time, tiered billing. Pricing changes in real time across multiple models and usage tiers; DeepSeek’s peak rate is 2x its off-peak rate. Airwallex, as the party issuing the bill, calculates the charge in real time and removes the reconciliation step. The rise of large models “is really a story of the past 18-24 months,” giving Airwallex a late-mover advantage.
• 2 contrarian decisions in Airwallex’s history created today’s moat: build a global network from day 1 rather than a regional one, and rebuild a complete global bank in the cloud rather than merely moving funds. “If you built a Southeast Asian payments business, you might have been profitable long ago,” but regional players ultimately sold to Visa or PayPal. “Only moving funds is a relatively poor business.” At the Series B, Stripe proposed acquiring the company, forcing a decision over whether to sell.
• Airwallex’s stance toward ChatGPT Finance is strategic wait-and-see. ChatGPT wants to connect with fintech companies, “but who do you see actually connecting with it? No one, for now.” The agentic-commerce analogy is Amazon and Walmart opting out because data is their most valuable asset. The core contest is traffic versus data: “If the underlying model becomes a generic commodity, why would I put the super-entry point somewhere else instead of building one myself?”
• The $100B roadmap is straightforward math: grow the customer base from 200,000-plus to 1 million and use intelligent finance to lift annual revenue per customer to $10,000. That supports a $10B revenue base and, in turn, a $100B-valued intelligent-finance platform.

Deep dive

1. Real-Time Tiered Billing: A New Business Fueled by the Global Expansion of Large Models

• 吴恺 contrasts traditional software billing—“set a plan at the beginning of each month, then send a bill every month”—with the “highly real-time, highly dynamic” economics of large models. Different models offer different performance at different rates; usage is priced in tiers, and rates keep changing. DeepSeek’s peak rate is 2x its off-peak rate, so “all of this logic needs to be built into a billing and payments platform like ours.”
• Koji asks why model companies do not calculate the charges themselves. The answer is process efficiency: Airwallex is the party that ultimately issues the bill and collects payment, while billing “is not their core capability.” If the model company calculates the charge and sends it over, “there is still reconciliation in the middle”—an unnecessary extra step. The trade-off is significant liability: underbilling means subsidizing the model company, while overbilling makes the end customer “extremely angry.”

2. The Starting Point Was a Melbourne Café: The Beans Were Cheap, SWIFT Was Not

• In Melbourne in 2015, CEO Jack and his co-founders were importing beans from South America and Africa for a café side business when they ran into 2 cross-border payment problems. Funds arrived slowly—李成’s pinyin matched the name of someone under sanctions, so SWIFT transfers were frequently held up—and the cost was punishing: “The beans themselves weren’t worth much, but 1 SWIFT transfer could cost you roughly $20-$50,” regardless of the amount sent.
• The technical insight came before the business insight. SWIFT “is essentially a messaging system”: it transmits information, after which banks move the funds, leaving the money flow separate from the information flow and making the system highly inefficient. As cloud computing matured and real-time, low-value clearing networks emerged—spurred in part by Chinese payment tools such as Alipay—the problem became one of connecting those networks to move money end to end in real time. Jack came from investment-banking FX algorithm engineering; another co-founder, 西京, was a serial entrepreneur who had worked at Google X. Wise and Revolut began around the same time.

3. The First Lesson in Vision: Go Global on Day 1, Even if the First 3-4 Years Are Loss-Making

• The first major choice was clear: “If you built a Southeast Asian payments business, you might have been profitable long ago, but we wanted to build a global network.” For the first 3-4 years, the company focused on bank integrations, local licenses and clearing-network access; it was only in 2018 that it “just started to have a little bit of revenue.” Today, Airwallex supports real-time settlement in more than 90 countries.
• 吴恺 is unequivocal about what happened to the regional players that stopped expanding their networks too early and rushed to monetize applications on top. “Their outcomes were all very similar: they were all sold to giants”—some to Visa, others to PayPal. “You monetize quickly, but the road ahead doesn’t go very far.”
• The strategic choice “only requires doing the math”: the largest payment rails are transcontinental. Execution is the hard part. Delayed profitability means needing strong investor-relations and fundraising capabilities, as well as a team willing to stay the course.

4. The Second Lesson in Vision: Transfers Are a Bad Business; Rebuild a Global Bank in the Cloud

• The second contrarian call was that “only moving funds is a relatively poor business.” It sits too far down the stack and can be used directly by only a limited set of B2B customers. To generate B2B traffic, “the most fundamental thing is to go back and build a global bank product”: corporate cards and wallet integrations on the payment side, acquiring gateways on the collection side, and the full end-to-end stack.
• Investors at the time thought the plan was “extremely crazy,” arguing that a startup could not build so many products at once. 吴恺’s answer was that the team was “hardworking and highly execution-oriented,” and ultimately built them all: accounts, FX payments, cards, acquiring, high-yield accounts and credit products. PMF on the global-banking track did not arrive until 2021. From 2018 to 2021, the company could only serve payment use cases—education platforms paying teachers worldwide, sharing-economy platforms paying landlords, and international students paying overseas tuition.
• His first-principles response to the startup doctrine of “go deep in 1 vertical before expanding” is customer need. Each product line was still “cut, cut, cut, cut, cut” down to the minimum usable state. But “when you cut away some of the capabilities central to the main workflow, your customers can’t use the product. You can’t make that compromise.”

5. Against Classic Banks: Coverage, APIs and the Line AI Is Erasing

• Airwallex’s 2 core selling points are coverage and experience. “There probably isn’t a bank anywhere in the world whose entire network coverage can match ours.” Banks build networks through their own branches, at high cost and low efficiency. On the experience side, Airwallex spans web, mobile and APIs—capabilities banks largely lack, given that their core systems “may have been built in the 1980s.”
• ERP integration is the clearest example. A company may process hundreds of invoices a month; through a bank, each one has to be entered, confirmed and paid individually. After Airwallex’s standardized integration, “the customer imports invoices from its ERP into our system, selects the invoices it wants to pay, and the money goes out with 1 click.”
• The line with Mercury is also beginning to disappear. Mercury is not API-first and serves very small businesses and SMEs. But AI changes the dividing line: “A problem that previously required a team of engineers may now be solvable by 1 finance operations person.” Financial automation for SMBs is becoming broadly accessible.

6. Kai and Agent OS: Behind the Interface Shift, Some Software No Longer Needs to Be Built

• Kai is a web-native conversational AI assistant covering finance-policy analysis, optimization of available cash flows, expense controls and workflow generation, and analysis of incoming-payment errors. Roughly 1 month after launch, “retention has been quite good overall”; once customers start using it, “stickiness is very strong,” and it is difficult to imagine them returning to the old way of clicking through everything. Agent OS gives customer-built agents a command line and an API-based MCP interface to call Airwallex’s full underlying capability set.
• For financial institutions with technical capacity, an AI assistant is “basically a must-have.” Mobile banking can expose hundreds or thousands of functions, and “you have no idea where they are; sometimes you don’t even know what the function is called.”
• The deeper shift is that software that was never built may no longer need to be built. “An agent doesn’t need to use software. Once you let an agent understand the logic of how to do something, it can do it itself.” Reconciliation and revenue recognition—processes that previously had to be performed one by one and consumed significant time—do not need to be developed as standalone software. “The customer may not care about the process anymore; they only need the result.”

7. The Verdict on Financial SaaS: 10 Software Products Become 1 Entry Point

• Looking 5 years out, expense management, accounting, revenue recognition, billing and other categories—“10 different finance software products”—could all be replaced by 1 large entry point backed by a set of agents. “5 finance employees using 10 different software products could become 3 finance employees using 1 agent system.”
• When the host points out that this means 吴恺 is extremely bearish on the future of financial SaaS, he does not deny it: “They need to change.” Data fragmentation was not a major problem in the past because other SaaS products could not deliver those capabilities anyway. In the AI era, an expense-management tool without other data has sharply limited AI capability; all its automation is fragmented, and “its AI is relatively dumb.” The conclusion is blunt: “They will need to spend a great deal of effort to change.”

8. Tzero, Airy and the M&A Logic: In the AI Era, Buy Only Data and Talent

• Tzero, as pronounced on the show, is an AI-native finance platform that gives a company “a mini-CFO from day 0”: bookkeeping, reconciliation, financial-statement generation, and cash-flow and income-expense forecasting. It launched first in North America and has received “a very good response in Silicon Valley.” Airy began with one-click checkout and ultimately aims to become a consumer-facing intelligent-agent wallet—one that can hold a balance, let an agent make purchasing decisions and settle transactions, and eventually support agent-to-agent payments.
• The logic behind acquiring Leap Fin, which focuses on revenue recognition, and OpenPay has only 2 parts: data and talent. Large amounts of proprietary data are not available on the public internet. Manufacturing has its own AR and AP patterns; software has accrual and revenue-recognition data; logistics services have another set of operating realities. “You simply can’t find this data online.” SaaS companies hold real customer operating examples that can help an agent build revenue-recognition capabilities “very quickly.”
• The pricing of talent has changed. AI amplifies founders’ capabilities: “A product manager who previously could only write documents can now produce a prototype.” As a result, “the premium we are willing to pay for this talent is much, much higher than before.”

9. Whether to Connect to ChatGPT: The Strategic Contest Over the Super-Entry Point

• Koji describes being pulled by a ChatGPT Finance pop-up into linking a bank account and asks whether Airwallex might one day also be operated by ChatGPT. 吴恺 is unusually candid: “I’m also trying to sidestep this question a little because I don’t have a particularly clear answer.” He starts with the current reality: ChatGPT very much wants to connect with fintech companies, “but who do you see actually connecting with it? No one, for now.”
• He uses agentic commerce as the analogy: the major e-commerce platforms will not play along. Amazon and Walmart treat their data as their most valuable asset, while companies such as Shopify are more willing to cooperate because they do not own the traffic. The central strategic question is: “If the underlying model becomes a generic commodity, why would I build a super-entry point myself and put the entry point somewhere else?” AGI adds another variable. Today, 1B people use ChatGPT for 20 minutes a day; after it becomes AGI, “I may use it for 1.5 or 2 hours every day and actually have it do things for me. That is completely different.” For now, “many people will be watching.”
• The internal trade-off is straightforward: “Our moat is our data, but behind us we also need its traffic, because intelligence itself is not the key point.” Airwallex serves several hundred thousand customers today; adding traffic could take it to several million or tens of millions almost immediately. “Intelligence is not a consideration here; traffic and the business are.” The strategy is to move forward and keep watching.

10. The Hardest Part Is Resisting Short-Term Temptation: Turning Down Stripe and Staying on a 6-Year-Old Model

• The hardest decisions are not about execution but choice: “Should we monetize and go public earlier? At 1 point, we even had the opportunity to sell the company.” Stripe wanted to acquire Airwallex during the Series B, leaving the company to decide whether to sell. “These decisions may be the hard ones: resisting the temptation of short-term benefits.”
• 吴恺 spent more than 3 years as CFO. In 2018, an investor dismissed the financial model he presented as “surely far too outrageous.” Revenue was “almost zero” at the time, yet the model ultimately projected revenue reaching the $1B-plus range. Looking back at 2018-2024, “we basically followed the model’s forecast, with the growth off by 1.5 quarters.” He also gives luck its due: “Looking back now at the model I put forward, I was definitely pretty lucky.” The method was simply vision-driven: identify the customers, size the transaction pool, and estimate how much revenue could be collected.
• The current valuation math starts with $1.3B in ARR growing 90%. “This PS multiple is not high—not high at all.” Industry gross margins are generally around 70%, “roughly the same as those of the top technology companies.” The 2 drivers behind transaction volume growing to more than 2x are new-market penetration—“we are still entering some new markets, but fewer and fewer”—and product layering. Per-customer value on the platform is growing at more than 20%, approaching 30% per year.

11. The Globalization Playbook: Locals Run Local Markets; GDP Sets Priorities

• The rule of localization is to hire a local GM. The process is “fail repeatedly, try repeatedly, fail repeatedly, then try again.” Early on, the company faced adverse selection: without brand recognition, the candidates it attracted varied widely in ability. The most vivid counterexample was a US commercial lead who said he lived in San Jose and could not make a meeting because of an emergency. 吴恺 and Jack stayed 2-3 extra days, but he never appeared. “Later he told us he was actually in Miami, vacationing with his family.” “We immediately decided to let him go.” Even after that internal friction, Airwallex still would not parachute someone in from headquarters to run the US.
• 吴恺 has considered why Huawei and Japanese and Korean companies can parachute in large numbers of executives while Airwallex cannot. Manufacturing companies build factories locally, where headquarters’ factory-building experience matters more than local culture. In fintech, the product is built entirely by headquarters; the local team’s core jobs are compliance and marketing, so “understanding local culture and rules matters much more than understanding the headquarters’ product.” Parachuting also creates a talent ceiling: the people he can attract by going to the UK are not the same as those an outstanding UK-hired GM can attract.
• Market prioritization was initially a 2x2 matrix of scale versus difficulty. It was later reduced to 1 rule: “A market that is both large and easy to enter doesn’t exist.” So Airwallex looks only at market size and ranks markets by GDP: “We first have to enter the 10 countries with the largest GDP globally. Once the top 10 are done, we go to 10-20; beyond 20, we assess case by case. Very simple.” The payoff is strategic discipline: “The thing we fear most is resources being high 1 moment and low the next.”
• The marketing mix varies by region. Asia has highly developed online economies, expensive traffic and relatively cheap labor, so field-led marketing can outperform digital acquisition through marketing staff. In Europe and the US, online costs are not as high relative to labor, so Airwallex built online growth teams when entering the UK, Europe and North America.

12. The Next Stop Is $100B: 1M Customers × $10,000

• In financing coverage, Jack 张 said the company was at “a crucial inflection point in the history of global finance.” 吴恺’s expansion is that traditional vertical financial software will be replaced under AI, creating an opening for a financial intelligent super-entry point. He rejects the view that super-intelligence exists only at the general-purpose layer: “We believe that in each major industry, there will still be a moat in specialized capabilities and non-public data.”
• The potential entrants are easy to identify. Stripe is “definitely building it”: it spent its first 10 years on payment gateways, but “over the past 3-4 years, its product lineup has expanded sharply,” because a super-entry point cannot be built around 1 product and intelligence is highly generalized. Revolut’s advantage is tens of millions of consumer users and the traffic they bring. “There are relatively few others—perhaps some regional players.”
• The $100B milestone is simple arithmetic: take the customer base from its current 200,000-plus scale to 1 million, upgrade the product from financial operations to intelligent finance that improves finance-team productivity, and generate roughly $10,000 in annual revenue per customer. “1 million customers and $10,000 of revenue per customer can support a $10B revenue base,” paving the way to a $100B intelligent-finance platform.