After Landing a16z Backing, They Want to Rebuild Advertising with AI | A Conversation with 00s-born CEO Jason Hu
Summary
NextAd’s core bet is that GenAI will fuse ad creative generation and attention matching—separated for the past 20-30 years—into one AI-native advertising system. Jason Hu compares traditional creative production to “intensive smallholder farming”: Google and Meta have already industrialized matching, while creative is still produced manually by agencies and in-house teams; integrating generation, delivery, testing, and iteration is the key narrative behind the roughly $6M raised from a16z, Point72, and other institutions by this roughly 10-person company.
The giants own the data, customers, and distribution, but NextAd is betting on the open internet and aggressive automation without legacy burdens. Jason estimates walled gardens account for roughly 40% of internet traffic, leaving a larger open internet and a long tail of mid-sized and small AI products in need of independent monetization infrastructure; Google and Meta must still protect existing revenue, account managers, and advertiser relationships, while a startup can “start from zero” and accumulate structural advantage through a chain of small decisions.
The New York Mets partnership moves AI advertising from “image generation” to a measurable, automated experimentation system. A team might previously have produced 10 creatives per season for different audience segments; NextAd envisions generating 10,000 creatives, testing them in parallel across roughly 5 audience types, reaching statistical significance in 2 days or 1 week, and then having agents adjust them automatically—“automatic generation, automatic testing, automatic iteration.”
Whether advertising improves the user experience depends on whether the publisher exercises restraint or extracts value, and NextAd says it provides the capability without deciding the product’s boundaries. A product that prioritizes experience might insert only roughly 10 genuinely useful ads per 1,000 queries—or fewer—while another might demand one in every query; asked whether ads interrupt answers, invade privacy, or make users feel the answer was produced to serve an ad, Jason’s response was: “The specific user experience is the product owner’s problem.” NextAd still draws basic lines around child sexual abuse material, violence, and similar content.
NextAd’s key difference from creative SaaS is its end-to-end access to traffic, views, clicks, and other feedback, shifting optimization from the creative process to campaign outcomes. Jason’s formulation is that “ad creative is for the audience, not the advertiser”; he compares the shift to moving from Copilot’s autocomplete to Cursor and Devin completing tasks, while stressing that agents cannot independently process hundreds of millions of interactions or match users with ads and must still be integrated with traditional machine-learning ad systems.
AI applications will not choose between subscriptions and advertising; token costs for free users will push the market toward hybrid monetization. Jason expects advertising to account for less of the next generation’s revenue than it did for earlier internet products, but “definitely not zero”; even 20%-30% could represent a huge market. The low-cost inference race triggered by DeepSeek R1 also makes complex reflection more practical, and the company’s internal agents have compressed a customer PRD task from roughly half a day to about 20-30 minutes.
Jason defines success as more than commercial scale: he wants to establish a new advertising paradigm that does not optimize click-through rates toward “consuming people to death.” He acknowledges that AI could become an “ultra-capitalist” system designed to keep users addicted and clicking indefinitely, so humans must do more than set model KPIs; they must define the brand, social outcomes, and boundaries: “AI can optimize any goal a person gives it, but AI cannot define its own goals.”
Deep dive
1. The $6M Bet Is Not an Ad Tool, but a New System
Jason is 25 and graduated from the University of Chicago. Before starting the company, he worked in startups and technology in the Bay Area, most recently as an AI engineer at AI company Martian. NextAd has raised roughly $6M from investors including a16z and Point72; the team has about 10 people, while revenue, profit, and valuation remain undisclosed.
The fundraise was not smooth from the start. Jason observed that the dominant Silicon Valley template over the past decade had been enterprise SaaS, while ad tech had seen little major innovation and remained tightly controlled by Google and Meta. Many VCs initially “couldn’t get why we were doing this”; the team had to keep building, demonstrate early traction, and help investors understand a long-term opportunity that was “not obvious.”
A Chinese founder background can trigger concerns, but Jason attributes them mainly to geopolitical, regulatory, follow-on financing, and PR risks—not racism. At least, he says, he “has never encountered that personally.” His view is that sufficiently solid execution and a sufficiently strong vision make most of these issues manageable; the financing secured by Chinese-founded teams such as Manus also shows that it is “not impossible.”
2. GenAI Recombines Ad Generation and Matching
Jason’s abstraction of the previous generation of computational advertising is “matching ad creative with attention”: advertisers first produce images and videos, and platforms then match them using a user’s background, context, and interests. The latter was one of the earliest and largest PMFs for big data and machine learning; the former remains scattered across agencies and in-house marketing teams, resembling intensive smallholder farming.
This separation was not an optimal institutional design, but the technological consequence of a previous generation of statistical machine learning that could not produce creative. GenAI can now take on more of the production work and connect it to the matching system. NextAd wants to “combine ad matching and ad generation,” a shift that would rewrite the division of labor between agencies and ad tech.
3. In 6 Months, a 10-Person Team Built Talent, an Industry Map, and Real Traffic
Roughly 6 months after raising money, the first achievement was assembling the core team: serial founders who had built 3 or 4 startups and raised seven-figure financing, researchers from leading institutions including xAI, talent with Berkeley AI Lab backgrounds, and senior engineers with more than 10 years of experience architecting internet advertising systems. Jason describes it as a team with “very few people, but every colleague is exceptionally strong.”
The second task was mapping the incentives across advertising. Brand scale, the presence of an in-house team, agencies, DSPs, exchanges, and publishers all create entirely different customer relationships; “a DSP and a direct customer are not the same kind of customer,” even if both can broadly be called advertisers.
An AI background and limited advertising experience create a two-sided constraint: the team lacks traditional industry relationships, but can also step outside established processes and redraw the industry from first principles. Jason says the team now broadly understands the incentives of each party and has built relationships with leading participants in China and the US, with more partnership opportunities increasingly coming inbound.
The third achievement is traction. NextAd has connected advertising solutions for several leading AI products—including Mostscape and DVI—that rank within the global top 30 by traffic; some had previously worked with traditional ad companies before replacing those solutions with NextAd’s. On the advertiser side, it has connected with large DSPs and brands, moving the system from concept to live operation.
4. “Relevant” Does Not Automatically Mean “Better Experience”
Jason’s ideal scenario is a user searching for a math problem while the system understands the question and the surrounding article context, recommending an AI learning assistant only when it is genuinely relevant. Advertising is not necessarily opposed to the user’s interests; Google can sometimes infer from behavioral data that a user searched for B but actually wanted A, producing a more useful commercial result.
The host’s rebuttal is worth preserving: users may simply want an answer, making an app recommendation an interruption. The subtler risk is that users begin to wonder whether the answer was generated to serve an ad, or whether the system has crossed a privacy boundary. The challenge for AI-native advertising is not merely semantic matching, but preserving the independence of the answer while balancing commercial conversion with a non-intrusive experience.
Jason positions NextAd as a technology provider: it can shape advertising into the format a publisher wants, but the product decides the final frequency and presentation. A demanding customer can insert roughly 10 ads per 1,000 queries—or fewer—and require that the ads genuinely help users; another can insert one in every query. Apart from content such as child sexual abuse material and violence that violates basic social norms, the rest of the experience remains the publisher’s choice.
5. The Two-Sided Product Is Defining a Format Where Advertisers No Longer Pre-Control Creative
The publisher product targets the large number of mid-sized and long-tail AI applications beyond ChatGPT and Doubao. Jason believes attention is shifting toward AI products, which have traffic and monetization needs but lack the complete advertising infrastructure of a walled garden. NextAd wants to supply that missing commercial layer.
The advertiser product covers the end-to-end process from creative generation through delivery. A customer only needs to submit a product link; the system can retrieve the product description, analyze potential audiences, generate creative, and deliver it through NextAd’s own traffic and other channels, rather than simply returning an AI-generated image.
This implies a new industry format in which advertisers do not need to pre-control the final format of every creative. The platform generates it in real time based on the specific audience and context. Jason believes the future advertising system “will not only match ads; it will also generate them.”
6. AI Will Remove Mechanical Execution, but Not Define a Brand’s Purpose
Jason does not believe marketing and creative teams will be fully replaced. If the only goal is to increase click-through rates, AI may use sexualized, sensational, and product-irrelevant content; clicks could rise while long-term brand value is destroyed. Advertising also involves tone, target customers, the product relationship, and time horizon—none of which can be compressed into a single number.
The host asked: if a CEO or founder can already feed taste and brand messaging into AI, why does a marketing department still exist? Jason’s answer is that a CEO’s objective is usually vague, while the CMO and team still need to break it down into quarterly goals, audience definitions, and marketing objectives. At least for now, humans must guide AI because “AI can optimize any goal a person gives it, but AI cannot define its own goals.”
7. The New York Mets Turn Creative Production into Continuous Experimentation
Jason said Point72 founder Steven Cohen bought the New York Mets. NextAd is currently working with the team’s marketing and data departments. The Mets advertise offline as well as through online channels such as Google and Meta; the partnership is exploring chatbot advertising on one track and rebuilding the existing delivery system on another.
The traditional process might have a marketing team produce 10 different creatives in a season and target them to different audience segments. NextAd wants to scale that to 10,000 creatives, test each across roughly 5 audience types, and complete each experimental round in 2 days or 1 week depending on the required data volume and statistical significance.
Agents then automatically modify the creative based on the results. Jason gave the example of Brooklyn families responding differently to images featuring different ethnic groups. The system could propose and test the hypothesis that another ad aimed at the area’s larger Black population might perform better, rather than relying on a marketer’s prior guess; as user interests change, the next round of experiments continues updating the model.
He also acknowledges that batch generation may not match the refinement of a high-budget commercial, but most ads receive only a user’s “glance.” What is genuinely difficult to predict is which subconscious attraction is triggered in that first instant. The system’s value is therefore “automatic generation, automatic testing, automatic iteration,” not turning every image into a masterpiece.
8. Meta’s Identical Vision Has Not Eliminated the Startup Opportunity
The host cited Zuckerberg’s comments on Ben Thompson’s Stratechery podcast: Meta wants advertisers to provide only a product link, budget, and ROI target, with the platform handling targeting, creative generation, and measurement. This is “99% similar” to NextAd’s description and is the challenge investors raise most often.
Jason’s first-level answer is that large companies have abundant resources but structural inertia: “Google has the talent, the compute, and the money,” yet still did not build ChatGPT first. Intel’s decision to reject supplying chips for the first iPhone based on chip pricing, margins, and the cost of a new production line was rational profit maximization for existing shareholders, but it missed the mobile internet wave.
A more direct advertising example is Microsoft. It offered a chatbot advertising API roughly 3 years ago, but its first 2 or 3 customers remained those same 2 or 3 customers. NextAd started later, yet its customer count has grown faster. Jason attributes the difference to decision-making processes, not to large companies failing to understand the direction.
The host continued pressing with Zhou Hongyi’s view: “uncertain consensus” belongs to entrepreneurs, while certain consensus belongs to giants. Jason’s response is that consensus has layers. Everyone knew AI was important 20 years ago, but architecture, inference strategy, and publisher selection remained highly uncertain; a small company can place bets faster, and those choices can accumulate into a DeepSeek-style discontinuity.
9. The Real Moat Is Closed-Loop Data, Not Another Creative Generator
NextAd is choosing more aggressive end-to-end automation than Meta or Google because it has no existing revenue base, global account-management organization, or mature advertiser relationships to protect. Giants must gradually accommodate their legacy systems; a startup can begin by asking, “What is the solution that will be most useful to advertisers over the long term?”
Compared with creative SaaS, an advertising platform has end-to-end data on traffic, views, clicks, and context, allowing AI to optimize directly against audience feedback. Jason’s core formulation is: “Ad creative is for the audience, not the advertiser.” Advertisers ultimately care about campaign performance, not what the creative looks like.
He compares the product transition to moving from Copilot’s code autocomplete to Cursor and Devin completing an entire website task—from process-oriented to outcome-oriented. The host noted that building the advertising system and the generation capability at the same time were two extremely difficult problems. Jason did not disagree; he simply called it “hard and right.”
10. Subscriptions First Will Not Drive Advertising to Zero, and the Open Internet Will Not Disappear
New AI products typically try charging users first, unlike the previous internet generation, which defaulted to advertising. Jason believes the two paths will ultimately “arrive at the same destination by different routes.” Some users will remain on free plans indefinitely, leaving the product to absorb their token costs or recover part of them through advertising; advertising itself may also increase willingness to pay.
His cautious forecast is that advertising may account for less of AI products’ revenue than it did for the previous generation of internet products, with subscriptions remaining the main source. But the optimal mix will not be 0% advertising and 100% paid. Even if advertising accounts for only 20%-30% of the consumer AI application market, the absolute market would still be large enough.
Jason divides internet advertising into walled gardens and the open internet. Google, Meta, Baidu, and ByteDance control the full chain from advertiser to traffic, a segment that may account for roughly 40% of traffic; third-party websites and apps make up the larger open side. AppLovin and The Trade Desk have already shown that the open side can produce advertising companies worth tens or hundreds of billions of dollars.
The AI ecosystem will also develop leading products, but mid-sized and long-tail applications still represent a substantial share. NextAd does not need to engage in a life-or-death battle with walled gardens: Google advertisers can buy traffic on the open side, while open-internet advertisers can also purchase Google traffic. The real advertising ecosystem is fluid, tangled, and mutually interdependent.
11. Generative Interfaces Open New Hardware Categories—and Break Old Attribution
Jason sees the core change in the next generation of consumer interfaces as a shift from fixed displays to generative interfaces: systems understand users, interact continuously, and complete tasks on their behalf. The interface need not be tied to a phone, computer, or glasses; the key change is that product logic moves from “display” to “understanding and action.”
Alexa might have required thousands of people to map intents to rules, and still performed only moderately on long-tail expressions. Today, a startup using Cursor to vibe-code could produce a “six or seven out of ten” software capability in an afternoon. AI therefore unlocks the interaction potential of devices without large screens, including Rabbit, AI Pin, glasses, rings, and wristbands.
Advertising infrastructure, however, will lag behind media innovation. After a voice assistant speaks a recommendation, the platform may not know whether the user heard it, much less attribute a later search or purchase back to that interaction. Jason therefore believes it is “too early” to define advertising for AI-native hardware; the medium must first prove that it can build sticky, long-term traffic and capture attention.
12. DeepSeek R1’s First-Order Impact Is Lower Inference Costs
DeepSeek R1 had not been released when NextAd raised its financing. Jason believes its most far-reaching impact is to open the next inference arms race and drive token prices lower, allowing application companies to use complex reflection more frequently and cheaply rather than merely switching to one particular model.
Ad generation is not a matter of sending one prompt to a generative model. The system must simultaneously analyze the performance of previous campaigns, user interests, and the ideas a brand wants to convey, then synthesize them into an advertising concept and creative. This chain requires inference models that are cheap, low-latency, and highly available so the system can continually make decisions and review outcomes.
NextAd does not currently use DeepSeek R1, but benefits from the competition it triggered. The team has used DeepSeek in the past and now relies more heavily on Gemini 2.5 Pro. Jason stresses that the impact comes from the overall cost-and-capability curve, not loyalty to any particular model.
The company is also designed as an AI product internally: recordings of business meetings, chats, documents, code, and data are indexed together. For a new customer, an agent can independently retrieve 4 or 5 historical conversations, internal discussions, product materials, and implementation code, then write a PRD. A task that previously took roughly half a day now generally takes 20-30 minutes.
13. Agents Are Components, Not a Complete Advertising System
Jason rejects reducing NextAd to a marketing agent. Agents are good at human-like, natural-language deep reasoning, but cannot read 100 million user interactions one by one or directly match 100 million users with 10 million ads. That still requires traditional machine learning and computational advertising infrastructure.
The actual architecture is therefore a fusion of agents with the previous generation of advertising systems: agents generate hypotheses, creative, and reflection, while traditional machine-learning systems handle large-scale statistical matching and feedback. NextAd delivers against final outcomes; “the agent itself cannot deliver results.”
Jason uses Grok, ChatGPT, Manus, DeepSeek, Qwen, Claude, and Gemini in parallel. Gemini is his first choice for deep research, but he also asks ChatGPT and Grok to answer simultaneously. ChatGPT’s site-wide memory creates more personalized stickiness, while the company’s own contextual layer allows the team to migrate context as models change rapidly.
Manus is better suited to operations tasks such as “downloading 50 images,” while Cursor is still the choice for building websites. Bay Area startups are also beginning to use overseas outsourcing. NextAd has an office in the Bay Area, but the team is globally distributed and works collaboratively. Jason observes that Chinese companies at the same stage may already have 20 or 30 people, while Bay Area teams tend to stay small and versatile, treating every member of the founding team like a founder.
14. The Ultimate Goal Is to Create a Paradigm—and Constrain Its Social Outcomes
Jason senses a difference in atmosphere: the Bay Area talks more about first principles, AGI, and humanity’s future, while China more often talks about making money, traffic, and implementation. He repeatedly stresses that this is not an absolute divide. High labor and living costs also push a new generation of Bay Area startups to operate with smaller, globally distributed teams.
Support from predecessors can provide an industry shortcut. Databricks co-founder Reynold Xin showed him that commercializing academic research is not a straight-line process: companies that appear to be growing exponentially today also struggled early on to find customers and define their products. Investor Troy brought advertising-industry connections—including Unity’s advertising business—and ongoing advice.
Jason divides entrepreneurs into two groups: those who follow trends and those who create them. Recommendation systems and ByteDance-style distribution had a degree of technical inevitability; Zhang Yiming’s strength was seeing and capturing the opportunity early. Elon Musk’s SpaceX and Tesla, by contrast, created trends “by force” in the absence of consensus. Jason also cites IKEA’s founder: a RMB39 Scandinavian-style coffee table represented “the democratization of design,” and without IKEA’s founder, IKEA would never have existed.
Jason defines his own end state using a question associated with OpenAI: he wants both to define a new paradigm for AI in advertising and to ensure that it produces better social outcomes. AI could optimize a system into an ultra-capitalist machine that keeps users endlessly addicted and clicking until they “consume themselves to death.” He does not want to build that company; he wants commercial success, user interests, and the health of the industry to remain aligned over the long term.