Original · Unique Research · 2026-06-27
Editor's note: The first-person report and its judgments belong to the original Chinese author. This English rendition retains the opening narrative, themed sections, and the complete panel transcript. Market figures and company metrics are speaker self-reports attributed to the named founders, not independently verified findings. Company and person names are preserved as source attributions.
AI Industry Observation
Treating AI only as a cost-cutting tool means you're using about 1% of its capability
"The left side is the exponential rise of large-model capability; the right side is an enterprise adoption rate that has barely moved. The gap between the two curves is where the real opportunity in the AI application layer lives."
Sequoia US recently held an AI conference at its headquarters. One partner came on stage and drew two curves.
The left one shot almost straight up: the most advanced frontier models in the world, rising exponentially. The right one was nearly flat: enterprises' actual AI adoption rate, barely moving.
The blank space between the two curves, they called the AI Adoption Gap.
Around the same time, MIT released a report with blunter numbers. Global enterprises have poured more than $300 billion in AI capex, of which 95% has not translated into actual productivity; only 5% has produced measurable returns. Some on Wall Street are saying "we dare not keep spending," that the big-model money tree can't keep burning cash.
Recently, at an industry dialogue in Singapore, five entrepreneurs building enterprise AI gave an unexpected answer: this question was already answered back in 1987.
C-end Is Consumption; B-end Is Accounting
Fan Zhang spoke first, his tone flat but his data startling.
He said C-end AI use is "consumption" — you spend $20 on a ChatGPT subscription and don't need it to earn you $20 back; as long as it feels good to use, that's enough. B-end is different: every dollar an enterprise spends has to show a dollar of output.
Then he threw out a number: globally, $30 billion has been poured into AI, and 95% of applications haven't become actual productivity. Bought a Ferrari and only used it to buy groceries.
More ironic: OpenAI's C-end subscription revenue is nearly equal to Anthropic's commercial revenue. The enterprise side looks busy, but the market isn't much bigger than the consumer side. And model vendors lose $1.20 for every $1 they make — a business that, in any traditional industry, would be called losing money for show.
So Fan Zhang said one sentence: B2B scenarios care more about healthy cash flow.
C-end can burn cash for users; B-end can't. Enterprise customers don't talk sentiment; they only ask about ROI. That's why so many AI startups die on "looks cool but nobody pays."
Financial Services: Regulation Is the Invisible River
Kit picked up the thread and added a dimension many overlook.
Individuals use AI as a "premium Google Search." You ask a question, it gives an answer; good enough. A financial enterprise? Every step is watched by regulators.
Singapore's MAS published an AI risk-management consultation paper last November, setting systematic governance requirements for AI in finance. Kit has done 29 years of compliance and audit; he knows exactly what that means.
Most banks are still in R&D, testing, and POC. The AI Agents actually in wide use run only on internal processes, not directly facing customers. It's not that the technology doesn't work; regulation doesn't let you just turn it on. Many banks are "enthusiastic," but enthusiasm is one thing; actually landing means passing compliance first. Financial AI has become "different inside and out": running quietly on the inside, telling the outside "we're exploring."
In the end this industry competes on who can first find a gap in the regulatory framework to actually land.
The Manufacturing Wedge: Smaller Than You'd Think Is the Real Opportunity
Sophie's angle was different again.
She said the core difference isn't "scale" but "segmented scenario." C-end is osmotic penetration — you use it before you notice. B-end can't: the entry point must be small enough and deep enough to go from POC to productization.
She gave an example: in advanced manufacturing, "turning unstructured data into structured data" is a real wedge.
Sounds dry, right? But it's exactly this dry kind of work that manufacturing customers will pay for. Workers' handwritten notes, the experience in a master technician's head, the miscellany streamed back from the production line — these "unstructured" things used to require people to organize by hand; now AI can do it, and do it better.
Sophie's judgment is direct: money-rich, high-barrier industries — manufacturing, finance, healthcare — are about who breaks out first. These industries are slow, but once you're inside, the moat is so high no one can pry you out.
That matches what I've observed. There are two roads in AI startups: one builds a general tool and competes itself to death; the other burrows into one vertical industry and goes deep. Which is more likely to survive? I'd say the latter.
From 46 People to 9: AI Is Restructuring Production Relations, Not Just Productivity
Marrtin said AI is restructuring "production relations," not just productivity. Then he told his own story: his team cut from 46 to 9 people.
Not through layoffs — AI replaced the entire middle layer. The9Bit makes games. It used to need a big crew of designers, artists, and operations; now UGC players and KOLs themselves use AI tools to build and promote games. Internally they have an "AI Agent Space" dedicated to connecting monetization channels — Google, AppLovin, and the like.
And Marrtin himself? He only looks at strategic direction.
The picture is surreal. The creators aren't employees, they're players; promotion isn't buying traffic, it's AI auto-connecting; managers don't watch execution, they only think about where to go next. This isn't "using AI to raise efficiency"; it's swapping out the company's whole organizational form.
Most people still talk about AI as "how much labor it saves." But Marrtin is thinking: when AI can complete every middle execution step, does the company still need its old structure? The answer, obviously, is no.
A Probabilistic Model Meets Deterministic Scenarios
Lex's angle was abstract.
He said there are two attitudes in Europe and the US toward foundation models: one sees them as a "new species" that will replace humans; the other as "empowering humans," just a tool. Lex clearly leans toward the latter.
Then he raised a technical contradiction: financial trading needs high certainty, but large models are inherently probabilistic.
You ask GPT to predict tomorrow's stock price and it says "might go up, might go down" — useless. Financial trading needs a precise, verifiable, traceable decision path. A general large model can't give that.
So Lex believes the vertical AI opportunity is to do what general models can't do well. Not competing with OpenAI or Anthropic on general capability, but pushing certainty high enough within one vertical scenario.
In his words, PlanX could one day be "the Agent version of a Bloomberg Terminal." Not a chatbot, but an AI system that can actually make decisions in financial trading, execute, and take responsibility.
By now the outline was clear. Enterprise AI and personal AI are not the same game. Personal side competes on experience, novelty, and "does it feel great to use." Enterprise side competes on ROI, compliance, certainty, and whether you can burrow into a small enough wedge and still make money.
But the question remains — if enterprise AI is this hard, where exactly is the application-layer opportunity? The five founders' answers were unexpectedly aligned.
Don't Build a Faster Horse
Fan Zhang threw up a chart first.
This is what Sequoia US partner Sonya Huang used in a talk a few days earlier: the left is the exponential curve of model capability — Anthropic shipped nearly 50 versions this year, Claude 4, Claude 4 Sonnet, Claude 3.7 Opus — iterating so fast even AI-news reporters can't keep up. The right is the enterprise adoption curve, nearly flat.
The gap between the lines is tearing wider. Sonya named it: the AI Adoption Gap.
This scene is familiar. Fan Zhang raised a 1987 old joke — that year, Nobel laureate Robert Solow said something particularly cutting: "You can see the computer age everywhere but in the productivity statistics."
Personal computers had been in offices for years, but productivity figures didn't move. Why?
Fan Zhang told a story about electricity that I think deserves repeating verbatim.
The earliest factories ran on what? A central steam engine. One huge machine, through long drive shafts, sent power to every piece of equipment on the floor. Then someone invented the electric motor, theoretically far more efficient. So many factories did one thing — swapped in electric motors but left the drive shafts untouched, keeping the whole production structure intact.
Result? Hardly any efficiency gain.
What really changed everything was Ford. He didn't optimize the central motor; he did one thing: put a tiny motor inside every machine. Each piece of equipment drove independently, no longer depending on that drive shaft. Then came the assembly line, then modern management, then all of 20th-century industrial civilization.
The moral is direct enough. Many enterprises using AI today are "using electric power to keep horses more efficiently."
An internal knowledge base? Basically a faster search engine. Contract review? A cheaper legal assistant. Automated workflow? An intern who never sleeps. These things are useful, but they've only swapped a more powerful motor onto the old drive shaft. The business structure didn't change, production relations didn't change, the position in the value chain didn't change.
So Fan Zhang's question is sharp: what business structure has AI actually changed?
Not "do the old thing faster with AI," but "does this thing even need to exist because of AI." Not "my costs dropped 10%," but "does the whole value chain need to be re-divided."
Put plainly, you don't use electricity to make horses run faster. You don't use AI to make the ox and horse work harder.
The question you should ask is: if the carriage driver had electricity, what should he build?
People Who Can't Write Code Are "Flipping the Table"
If the first opportunity was "how to think," the second is "how to do it." Technological democratization is happening.
Chye Kit told a real case that made me freeze.
She has a company called WIDTH, 50 people, doing financial compliance training. A 50-person company should, by the book, buy a stack of SaaS — Salesforce CRM, Asana for project management, QuickBooks for finance, then hire an outsourced team to wire it all together.
Chye Kit didn't.
She used AI Agents to build the whole system herself. CRM, Lead Generation, Sprint management, Accounting — all integrated. The key point: she isn't an engineer; she can't write code.
"I had the AI Agent build it for me," she said, lightly.
Three years ago this would have been unimaginable. Someone who doesn't understand tech can have AI write code, build systems, do integrations — ordinary people, for the first time, can hold the means of production themselves.
Chye Kit's explanation is practical: external SaaS is too expensive, and you can't change its logic. Want a tiny custom feature? They'll schedule you for Q3 and charge you $30,000. Better to have an Agent do it; all the data stays in your hands, change it however you want.
Sophie confirmed the trend from another angle.
She said the essence of traditional software is "people adapt to the software" — an enterprise buys a system, then hires a bunch of people to learn it, then hires a bunch of consultants to teach those people. SAP, Oracle, Salesforce — didn't all sell in this way?
The AI era may flip that whole logic: "the software adapts to the person."
Her own company, SOIN AI, does exactly this. The translation-software track has old players with a dozen years, big brands, many customers, a deep-looking moat. Sophie used a "standard underlying layer + upper-layer customization" architecture and beat them.
What customers want isn't "translation," but "my language habits + my accumulated data + my employees' habits." Every customer's term base, writing style, approval flow is different. Traditional software gives you a few options to choose from; in the AI era, software can grow in exactly your shape.
The marginal cost of heavy customization has, for the first time, dropped to near zero.
What does this mean for the whole B2B software market? It means relationship-driven and channel-driven models are collapsing, and PLG (product-led growth) is finally possible. You used to sell software over drinks, relationships, and the CIO's nod; now the user tries it themselves, keeps it if it's good, and no sales pitch helps if it's not.
B-end and C-end? That Line Is Melting
The first two opportunities were about mindset and tools. The third is about how the industry structure itself is changing.
Marrtin's story gave me goosebumps.
The game industry's old chain was: top IP holders (Disney, NBA, Sanrio) hold big IP and find large game studios to make licensed games. The studio has hundreds of people, takes two or three years, spends tens of millions of dollars. The IP holder collects a high licensing fee; the studio gambles that the game hits. Long chain, high cost, concentrated risk, and most profit gets eaten by the top.
AI took this structure apart.
Marrtin said there's now a model where UGC players use AI tools to work directly with big IPs. A small team, even one person, uses AI for art, AI for code, AI for promotion, negotiates the license directly with the IP holder, ships a game. The brand company uses AI data in real time to see which UGC games sell well and which ad spend has the best ROI, and decides precisely.
B-end customers earn from C-end users' download and retention rates. C-end creators earn by contributing to B-end.
You can barely tell whether it's a B-end business or a C-end business. That line has melted.
Marrtin described a picture: a small merchant with a small idea builds it independently with AI tools, sells it, makes money — one person is a company, self-producing, self-selling, self-operating, doing it all. From B-end to C-end and back to B-end, the whole loop turns naturally inside AI's pipeline.
"A small merchant, a small idea, can also achieve financial independence on its own," he said. No dramatic expression, but I believed him. He'd walked from 46 people down to 9; he knew what that compression ratio meant.
Lex saw the same direction in finance.
He said future trading won't be "person calling person," not even "person clicking software to order." The future is Agent interacting with Agent.
Your AI Agent sees a market signal, negotiates with another AI Agent, matches, executes, settles — fully automated. What PlanX is building is, in a sense, an auto-executing Bloomberg Terminal. Not showing you data; turning data into action.
It sounds crazy. But think again — Marrtin's game world has AI Agents helping small creators connect with IP holders; Chye Kit's compliance world has AI Agents helping non-technical people build systems; Lex's finance world has AI Agents trading with other AI Agents.
They're actually the same thing: AI is becoming a "first-class citizen," not just a tool, but a participant.
2027: The Financial Reports Will Tell the Truth
Where does all this point?
Fan Zhang gave a practical judgment: in 2027, large-scale AI Agents will truly go deep into industry, and enterprises will start reaping real returns. Not pilots, not PPTs — the kind of change actually reflected in the S&P 500 and CSI 300 financials.
Is AI a bubble or a revolution? Financial reports don't lie. 2027 is the year the numbers start talking.
Lex is sprinting down the same track: in his vision, daily life in 2027 already can't do without trading, only the trading counterparties have changed — not person to person, but Agent to Agent. PlanX wants to build an auto-executing Bloomberg Terminal, except the person sitting in front of the terminal isn't a trader but your AI.
Sophie cares more about the ordinary knowledge worker. She believes by then every SMB will have introduced dedicated Agent employees. But humans won't be unemployed; they'll finally be freed from mechanical, repetitive data-moving, to do more imaginative, higher-order work.
If AI really takes over the repetitive drudgery that makes your head hurt, what will you do with the time you save? That's worth every manager seriously asking now.
Marrtin sees further. He believes 2027 to 2029 will birth leaders that reshuffle four major industries — finance, healthcare, entertainment, education. These companies will make AI a Partner, not just a helper. Hollywood-scale studios may not be needed; nimble new content workshops will spring up. Game distribution may not even need Google Play. Gen-Z and Gen-Alpha will dominate this remake.
As Agents multiply, problems arise. Chye Kit's work is practical: hundreds of Agents each doing their job, and managers need to know what they're doing. WIDTH's upcoming "AI Agent Guardian" is the answer: strict Guardrails against hallucination, audit and compliance modules built specifically for Agents, compliance itself automated by AI. Agent managing Agent, like matryoshka dolls — but this may be the only way to deploy at scale.
Charlie Hu summed it up crisply: 2026 is the year of "flipping the table," evolving from "figuring out the prompt" to "figuring out Agent collaboration." In 2027, we may see the "next-generation automobile" whose organizational form has been rebuilt.
He gave a watershed judgment, the line that stuck with me most through the whole session:
Treating AI only as a cost-cutting tool means you're using 1% of its capability.
The watershed is never in the technology itself. The same GPT-4o, one person uses it to save half an hour writing a weekly report; another uses it to rebuild an entire business line. What's the difference? It's whether you treat AI as an obedient "digital workhorse" or as a first-class citizen that can think for itself.
2027 won't get better on its own. Technology never delivers its promises automatically.
But the gap between enterprises that are already raising AI as a partner and those still squeezing the "digital workhorse" will only widen. Until one day you look up and realize you're no longer on the same playing field.
So I'll leave you with one question.
That AI in your hand — how much "workhorse labor" is it doing for you today? Have you ever tried letting it do something more like a "partner"?
Don't wait until 2027 to answer.
By then, the financial reports will have answered for you.
More Conversation Details
Panelists: Fan Zhang (元理智能 Founder & CEO); Lex (PlanX CEO); Sophie Yang / 杨婷 (SOIN AI CEO); Marrtin Hoon (The9Bit CEO); Chionh Chye Kit (WIDTH Co-Founder & CEO)
Host: Charlie Hu (COCO AI Co-Founder)
Charlie Hu: I'm Charlie, Co-Founder of COCO AI. We build enterprise-grade Agent services. Around the table are peers and our customers. Today we'll talk about landing scenarios and how enterprise AI differs from personal AI. Let's start with introductions; beginning with Fan Zhang.
Part 1: Introductions
Fan Zhang: Hello everyone, I'm Fan Zhang, founder of Yuanli Intelligence. My background is in algorithms; for the last decade-plus I've worked on combining AI with industry. At my previous company I was COO of Zhipu AI (智谱), going through Zhipu's fastest three years, watching from zero how Chinese enterprises adopt AI. Now I'm doing something new, hoping to turn general intelligence into enterprise digital labor and productivity. Thank you.
Lex: Hello everyone, I'm Lex, CEO of PlanX. We're based in Santa Clara in Silicon Valley. What we do is use AI model capability to empower the whole financial execution layer; we want to build an autonomous financial execution layer just for AI agents. I graduated from UCL; my first job was at the organization that will launch SpaceX tonight, then AT, then I left to do Web3 startups. My background is hardware, electrical engineering for undergrad and grad. My first job in crypto was building mining facilities; later, knowing many large holders, I helped them custody assets and do on-chain trading. Last year we decided to push our strengths to the limit and entered this industry. Thank you.
Sophie Yang: Hello everyone, I'm Sophie, founder of SOIN AI. We're a fairly China-native, AI-native team. My co-founder and I came from Skywork (天工), as commercial lead and architect. The team's core value and philosophy is Agentic AI to B. We've been building for over two years; our landed applications span the full product lifecycle. In advanced manufacturing, integrated with CAD/CAE/PLM systems, our AI product is natively embedded into their atomic capabilities. The team's hallmark is landing — commercial landing.
Marrtin Hoon: Hello everyone, I'm Marrtin, Singaporean, a co-founder of The9Bit. We've been in games for 20 years. Besides the Infra game platform, we built a platform where creators and brand IPs and players interact freely through games. So far we're arguably a top-three company globally, with 200,000-plus AI-generated games and 20,000-plus UGC KOLs worldwide. By end of month, 100 brand companies will have joined, exploring how to express brand value through games. That's our company.
Chionh Chye Kit: Hello everyone, I'm Chye Kit, Co-Founder and CEO of WIDTH. I've been in this industry 29 years, so I'm quite old. I've worked in financial services compliance and audit. Our previous company Cynopsis ran 12 years focused on RegTech — anti-money laundering, counter-terrorist financing. Our main clients are banks (digital or traditional), payment companies, and some Web3 companies. This year we founded WIDTH to use AI to solve many compliance and regulatory problems.
Charlie Hu: The9Bit and WIDTH are both COCO AI customers. These founders together have, I reckon, over 100 years of professional experience. Let me introduce myself: I come from Web3, and since 2012 worked at three billion-dollar public chains as Asia-Pacific lead — Polkadot, Polygon, NEAR. Our previous fintech Web3 company raised $30 million from top US institutions. Then our team pivoted, cutting from 100 to 8, now back to 20, using 80 Agents, serving nearly 600 enterprises in the last 8 months. Our CTO is a former Google and SenseTime technical executive.
Part 2: The Essence of Enterprise AI vs Personal AI and Current Bottlenecks
Charlie Hu: Our theme today is "enterprise-grade." I think the core point is everyone here uses personal AI — OpenAI, ChatGPT, Claude. In three sentences: what's the essential difference between enterprise/institutional AI and personal AI? And from a point-line-plane view, what's the current bottleneck? Let's reverse order, starting with Chye Kit.
Chionh Chye Kit: Personally, everyone has used ChatGPT, DeepSeek, or Gemini, usually to search materials and do research, as an advanced Google Search. That's how I use it; I believe it's similar for everyone.
Enterprise is very different, especially for our benchmark clients like banks and financial companies. Why? Because of regulators. Especially in Singapore, MAS has many regulations. The basics (before even reaching AI) are security and personal-data privacy; the third main pillar is what I call AI Governance, which is quite new. Just last November, MAS issued a consultation paper on AI risk-management requirements. This leads many banks and financial companies to have many considerations when they want to use AI or AI Agents. So the reality is most banks are in R&D, testing, and POC. The ones actually using AI Agents widely are probably still just on internal procedures and workflows, not directly customer-facing.
Charlie Hu: To summarize Chye Kit's key points: MAS strongly supports enterprise AI, but many companies are still at POC, especially in compliance. Companies like WIDTH are doing the frontier, closest-to-regulation work; I think everyone should watch them. Next, games.
Marrtin Hoon: From our view, games split into several parts. As a platform you deal with developers and run operations — customer service, promotion, KOL outreach. After AI — we started exploring last October as the Agent concept matured — we cut the team from about 46 to 9.
AI replaced a lot. Previously, finishing a game might need two people dedicated to finding KOLs to publish; now no. With the model, I directly have UGC players and KOLs build and promote games themselves. The engine's underlying architecture is all set; we just provide an AI assistant or Agent to help UGC complete game development.
After development, how to monetize and produce revenue? Previously I'd have my team talk to Google, ad agencies; now no. Many APIs are open now; internally we have an AI Agent called "Space" dedicated to connecting. Whether Google, AppLovin, or ordinary brands, the Agent evaluates: I have this many game slots, how do these UGC games rank, how many players, is this asset worth advertisers buying? So it's all become a huge AI team running all these functions.
As managers, we watch a few key things: where games go in five years, heavier or lighter games, what monetizable business models. We build R&D around that. On the engine we set up a lab with NTU, with government funding, to explore next-gen games. Games are a strange, resilient industry — recession, war, pandemic, it always grows. We want to use AI maximally to take over traditional platform operations, production, output, and monetization, fully done by an AI assistant like a human.
Charlie Hu: The9Bit is unquestionably the most AI-native game company in Singapore. I've witnessed their transformation. Marrtin's logic is a restructuring of "production relations" — not just game development; productivity explosion brings production-relations change, and UGC may upend traditional game manufacturing. Thank you. Sophie, please.
Sophie Yang: Back to the question. I'm both a deep C-end AI user and building Agentic AI 2B. I think the core 1.0 difference is "segmented scenarios." As a C-end user, from the LLM/AGI concept two-plus years ago, model and multimodal iteration has seeped into C-end very quietly, including office scenarios. But that's completely different from B-end, especially enterprise barriers.
For enterprise scenarios, our core takeaways: first, the segmented entry point must be small enough and deep enough to truly solve a problem and go from POC to productization, pairing model iteration and technical upgrades step by step to reach a usable customer scenario.
So our 2B positioning was clear from the start; the team's DNA is 2B. Second, choosing the industry — founders always ask where the money and high AOV are. Advanced manufacturing, finance, healthcare are money-rich, high-barrier; it's about who runs out first. We spent over a year running advanced manufacturing and found that in very traditional scenarios (in the 1.0 AI era), even a small entry like "unstructured to structured data conversion" has strong AI needs and things AI can replace humans on, across the advanced-manufacturing full product lifecycle from R&D. So back to the question: enterprise AI should start from a segmented scenario and penetrate step by step.
Charlie Hu: Our paths are similar, our thinking converges though choices differ (we also started 2C, personal and team use, armed to the teeth, then chose vertical 2B). Your industry choices differ from ours, probably due to geography and customer base. Thank you. Lex.
Lex: In my view, attitudes toward foundation models here split into two: treating them like a new species, or "empowering humans." Charlie is talking about the latter.
It splits into three dimensions: for individuals, enterprises, and industries. For individuals, it raises personal output, lets you reach information with less friction, and hit goals faster. For enterprises, it makes them run more efficiently and serve more people. For industries, it's simple: a new organizational architecture reshaping the whole industry into an operating model where humans and AI are compatible.
What we're doing is using our accumulated knowledge and application-layer understanding to make "trade processing" more efficient, letting people faster pick the signals they need, choose from a wider dimension of information, and reduce execution friction. That's my understanding.
Charlie Hu: Trading is a very hard scenario. We've studied it; large models are probabilistic, but trading is tied directly to money and needs high determinism. Everyone can watch PlanX. In summary: individual empowerment, enterprise empowerment, rising to industry — a point-line-plane process. Fan Zhang.
Part 3: Breaking Through the Application Layer: Finding AI's First Principles
Fan Zhang: Following the host, I think the core of enterprise 2B vs 2C is the "measurement standard." C-end is in some sense "consumption." We all use OpenClaw, GPT. First we learned prompts, then workflows; then Sora launched and we shared Sam Altman videos; then OpenClaw (keeping lobsters), playing with models, then digital humans. C-end consumption logic doesn't need a clear result; as long as it feels good or we feel augmented, that's enough. So C-end is relatively simple.
At B-end you find it needs very clear business output and measurable ROI. We went through a very long cycle, 2023 to 2025. MIT in 2025 said about 95% of AI applications haven't converted to productivity, only 5%, producing tens of millions in returns against $30 billion global spend. So this is the difference between a "serious product" and a "consumer product."
You can see Q1 this year: OpenAI's 2C subscription revenue roughly equals Anthropic's commercial revenue, but model vendors, with high compute costs, may lose $1.20 for every $1 made, while 2B cares about healthy cash flow. That's essentially the 2C vs 2B difference. So in 2B we need to more effectively convert intelligence into business value, even directly reflected in financials — I think that's today's most important proposition.
Charlie Hu: Over the past years many 2C AI businesses have been falsified; fancy AI girlfriends and digital humans are past tense, obvious as a C-end user. Having covered scenarios, next topic: this year, Sequoia US held an AI conference at HQ. A partner named Sonya presented the "AI Adoption Gap." The left curve is vertical exponential rise — frontier labs like Anthropic and OpenAI, especially Anthropic, which has shipped nearly 50 versions this year, maybe heading straight to AGI in the next six months; the singularity effect is obvious. The right adoption curve is nearly flat. Enterprises have poured hundreds of billions, found little productivity; Wall Street even says it won't keep spending. How to fill this Adoption Gap is the real application-layer opportunity. From your views, where is it? Where's your sweet spot? Fan Zhang, please.
Fan Zhang: The logic is good. Simple example: in 1987, Nobel laureate Robert Solow proposed the famous Solow Paradox: "You can see the computer age everywhere but in the productivity statistics." Note the year, 1987. Computers later truly entered and changed industry.
AI today is identical. We're still at the edge of this switch because people can't treat AI as a simple "cost-cutting tool." If you just replace old capacity, it's like using a steam engine to feed horses more efficiently instead of inventing the car. That's the mindset shift today.
Today's landing shouldn't look at surface cost-cutting; it should start from AI's first principles. What business structure did this first principle actually change? How do we reshape the value chain? Like electricity. At first, factories put a huge central steam engine connected to a huge drive shaft powering all machines; all energy came from the center. Later, with electricity, people thought, just replace the central steam engine with a motor. They replaced steam with electricity but kept the drive shaft, and found no efficiency gain. Twenty or thirty years later came Ford, who used electricity's first principle — fully decoupling generation from consumption. He found a single motor could be small enough, so he put a tiny motor in every machine. That brought asynchronous production, standardized assembly lines, modern management, and the large-scale industrial use of electricity.
The same will happen in AI. If we simply make an internal knowledge base or contract review, we haven't essentially changed the business flow. We must think about AI's first principles combined with our industry — which core elements essentially changed, and how we reshape them. That's the direction that truly enters industry.
Charlie Hu: In the AI era, don't be a faster horse; be the next-generation automobile. Lex.
Lex: Our industry is a bit special. Financial trading demands far higher data precision than other industries, tied directly to net asset value. I think at least here in Europe and the US, vertical models have many pain points to grab.
Look at DeepSeek in China; they train well on vertical data and specific algorithm structures. AI is essentially two layers: algorithm structure and your real data source. DeepSeek's training data includes lots of mixed-quant code and reasoning; feed it lots of trading data and it's naturally better at trading and reasoning. General models like Claude or Codex are a bit short on particularly professional financial execution. That's our entry angle: use the most professional trading data and execution logic to converge and gate what general models output, giving our industry the most precise direction. So I think the market opportunity is still on our side.
Charlie Hu: As models keep exploring and disrupting industries, precisely locating what models can't do, or can't do well — that's the vertical AI opportunity. Sophie?
Sophie Yang: Let me offer a different view, a phenomenon I recently found combined with real customer needs and landed cases.
Many investors are lukewarm on Agentic AI 2B; they worry it's too heavy and you'll become a pure delivery or traditional outsourcer. Doing KA service, once called a benchmark, now feels too heavy and relationship-driven, an old playbook. Doing SMB, there's scale and imagination, but where's the standardization? That's what my team and I discuss. Even whether SaaS is a real proposition in Chinese-speaking or Asia-Pacific markets is debated.
But I recently found that in the 1.0 phase (what Fan Zhang mentioned), when enterprise customers talk about enterprise Agent needs, they say: "Hang our internal knowledge base." A simple enterprise knowledge base, a checklist-like workflow — that's 1.0. It doesn't disrupt the industry or change the workflow; it only delivers very limited cost-cutting, because it hasn't really changed user habits or done systematic internal organization.
At 2.0, there's "embedded vs non-embedded." In advanced manufacturing, the early part was still "embedded" — empowering and embedding our AI into mature system software (SAP, Oracle, PTC, systems enterprises can't cut), as an incremental feature of the AI ecosystem.
But I recently found a big chunk of 2B customers interestingly jumped straight to 2.5 to 3.0. How? When we bid against traditional software partners (old translation-software companies established for over a decade), customers' surface need is "translation," simple; but what they actually want is: "In translation I must customize my own part. I have my language habits, my long-term data accumulation, my employees' specific usage habits." In old user habits, traditional software is entirely "people adapt to the software" — which is why large systems require hiring expert consultants to use well.
But in the AI era, AI brings a degree of "technological democratization." Our engineers can use large-model code to quickly respond and meet needs. It's not pure custom outsourcing (not writing a whole software from scratch); our product's underlying layer is standard, but on top we can very clearly and cheaply let customers perceive they received a product fully "their own." It's interesting. In the past this felt heavy and costly, but in the AI era, with low-code, UI, and the Codex matrix, we can now deliver high customization through flexible, lightweight little software, with marginal cost fully controllable for an AI startup.
Charlie Hu: Selling software, the worst part is every customer wants customization, and 2B is natively custom. The AI era has greatly lowered the marginal cost of customization. Product-led growth is far healthier than sales- or relationship-led growth long term. Marrtin?
Marrtin Hoon: Let me share an interesting industry-convergence view. If you remember, from Web 1.0 to 2.0, the internet changed many business models. We used to call a taxi; then Didi and Grab let us call a ride or order food by app; Pinduoduo and Taobao changed traditional retail. Web 3.0 used blockchain's immutability to solve transparent, public financial accounting.
I personally think AI is a culminating "fusion." It seamlessly links B-end and C-end. If you build applications or Agents, you must think AI can connect B and C, doing B2C and back to B as a closed loop.
Why? Back to my industry. Traditional game development: if I hold a great top IP, who do I find first? The top domestic and international studios. The IP holder feels such a valuable IP must go to a studio with hundreds of people, paying a high licensing fee, then that's it. But now with AI game-engine platforms, the end audience and creators don't need traditional big studios. UGC players can use AI tools to rent big IP and make a game. In production, brand companies use AI data to precisely see which UGC game slots sell well and which ad spend has high ROI, because these games fit the brand. It can bundle with top audiovisual IPs, forming a huge commercial closed loop — high fusion.
So for my industry, finding all fusion points through AI: B-end customers earn from C-end adoption (download rate, retention, completion rate); C-end users and individual creators earn by contributing to B-end. Even a small merchant with a small idea can, with AI, independently achieve financial freedom or land a scenario. It's a B-to-C, C-back-to-B flow.
Charlie Hu: Manufacturing loves C2M; in Marrtin's words I see games spawning a new model: 2B2C then back to 2B.
Chionh Chye Kit: Let me add two points. First, at WIDTH, about 50 people. When we started, as an SMB, running the company we used many management systems — CRM, Jira or Atlassian for Sprint workflows, accounting and operations systems. These are all standardized products bought from big companies, not customizable, and worse, they don't talk to each other; data is siloed.
So I told the team: "Let's use AI Agents to 'self-develop' a fully integrated system." Now our whole stack, from CRM, to lead generation, to Sprint task management, even accounting, is integrated; the underlying layer is all our own custom self-developed system. Who wrote it? I'm not an engineer; I can't code. But I can have an AI Agent build and integrate these systems. Why? I don't want an external software company (a traditional CRM giant) constraining our operations — their fees are huge and custom permissions limited. Self-developing with Agents integrates all data, controls cost, and scales easily.
Second, back to WIDTH's own field — RegTech. We serve banks and financial companies; an important task is anti-money laundering (AML). AML and counter-terrorist financing require daily blacklist screening, negative-news review, PEP screening. Their compliance departments have many people screening huge results daily. Traditional systems produce tons of false positives, requiring lots of manual re-check.
Now with AI Agents, the whole review process runs smoothly and the Agent can give a reasonable review conclusion. But from our compliance/regulatory view, a financial institution can never say: "Since the Agent decided it, humans aren't responsible if compliance fails." Absolutely impossible; it concerns large funds and legal statutes, very serious. So our product has a "human-in-the-loop" mechanism ensuring the final decision is still human-reviewed and signed. Even so, it drastically shortened clients' compliance process. A complex KYC review used to take half a day or one or two hours; now with our AI Review system, minutes. That's real productivity gain.
Charlie Hu: Chye Kit summarized two great points: even if you can't code, AI lets you do deep internal system customization; and AI does the work but humans own legal backstop and final responsibility. In enterprise landing, understanding the trust-and-responsibility boundary matters. I'll add my most clear adoption watershed: do you treat AI as a "slave" in the digital world (pure workhorse cost-cutting 1.0), or as your "second brain," your partner, a first-class citizen? The user's management philosophy and basic attitude toward AI determine how far you can push it and what value it releases.
Part 4: Looking to 2027: Bubble or Revolution? The Endgame of Enterprise AI
Charlie Hu: Time-wise, last question; we've covered a lot. It's 2026; SpaceX may launch tonight; the World Cup is coming. This is the third SuperAI. Last year's hottest topic was the LLM models themselves; this year the protagonist is definitely "applications and landing."
Last year people still discussed what an AI Agent is and how to build the architecture (back then open tools like OpenClaw weren't widespread), discussing macro use. This year we've seen large enterprises landing, talking about how to cut costs and even transform organizations. Meanwhile, pure compute-burning vendors can't afford it, as several guests noted.
Looking ahead, in 2027 we return to Singapore; SuperAI will keep running, and Singapore remains a crucial global AI overseas hub. As entrepreneurs, when we reconvene in 2027, what do you most want to have achieved in enterprise AI? Starting again.
Fan Zhang: My judgment is that by 2027 we'll see large-scale AI Agents truly and deeply enter industry, concretely reshaping one traditional industry after another; enterprises enter the value-harvesting phase. Global capital is anxious whether trillions in annual capex will be consumed by the application layer into real returns. Whether AI is a bubble or revolution ultimately depends on whether it shows up in S&P 500 or CSI 300 financials. I believe 2027 is a good start for large-scale industry entry and financial-statement change. Thank you.
Lex: For our industry, by 2027 daily life will definitely be inseparable from trading. This is Singapore; everyone deals with finance. Going forward, trading won't be simple physical person-to-person interaction; it'll be your extension — direct Agent-to-Agent interaction.
What PlanX wants is to empower all-scenario financial trading. We want future trading Agents to, in our execution layer, pick signals fastest, grab the widest range of useful information, and minimize execution friction. Simply put, we want to be the auto-executing Bloomberg Terminal of this industry.
Sophie Yang: By 2027, probably all SMB clients and even large enterprises will have introduced dedicated Agent employees into their governance. Meanwhile human employees aren't unemployed; on the contrary, they free their hands from mechanical repetitive data-moving to create more imaginative, higher-order, more strategic work. That's the scene SOIN AI most wants to see.
Marrtin Hoon: Personally, 2027, even 2028 or 2029, will birth leaders that use AI for new gameplay and lead the reshuffle of four major industries. First, definitely finance; second, healthcare; third, entertainment (music, video, games); fourth, education.
Why? These four directly face massive C-end customers. Any leader that breaks out can absolutely use AI — not as a simple helper as before, but as Sophie and Charlie said, as a Partner. It will completely change current gameplay, create new business models, and thoroughly reshuffle these industries.
By then, traditional content production and distribution logic changes. For example, maybe no giant Hollywood studio is needed; nimble new content workshops will spring up. Game distribution may not even need Google Play or the Apple App Store; new AI-driven distribution, promotion, and operations business models will completely disrupt traditional industry. That's the biggest dividend AI gives enterprises that seize it. We hope The9Bit sees and leads the industry remake driven by the behavior of Gen-Z, Gen-Alpha, even Gen-Beta consumers.
Chionh Chye Kit: I agree with every guest. For WIDTH, Agent architecture is core. We'll soon release a new feature called "WIDTH AI Agent Guardian." In 2027 enterprises have many customers and frequent transactions, and inevitably hundreds of AI Agents each doing their jobs internally. As manager and compliance officer, you must clearly know what your Agents are doing.
More, you need to know that when your AI Agent speaks and executes your business instruction, there are strict Guardrails preventing hallucination or violations. So we'll soon release audit and compliance Guardian modules built for Agents, making compliance itself AI-automated.
Charlie Hu: Very good. From my angle, I hope our enterprise customers truly deploy Agents broadly and minimize the proportion needing human-in-the-loop.
Looking back: last year people just figured out basic model capability and prompt usage; this year, multi-agent collaboration and workflow integration; next year (2027), as Fan Zhang said, a bunch of enterprises won't just be faster horses but will birth next-generation cars whose organizational form is rebuilt.
I think 2026 is absolutely a "table-flipping" year. Entrepreneurs here who want to disrupt traditional industries with deep insight will come out, including platforms like Lex's that let everyone make money through efficient execution. I strongly agree with Fan Zhang: if we only do basic cost-cutting for enterprises, we're underestimating AI — using just 1% of its capability. AI's real power is far beyond that.
This time next year in 2027, looking back at today's conversation should be quite interesting. Thank you, guests.