The Most Underestimated AI Buyer Is the Shop on the Corner

Alyssa Henry had just stepped down as longtime CEO of Square, inside Block, when she sat down with Sarah Guo and Elad Gil on *No Priors*. Before Square she was a vice president at AWS, the person Guo introduces as running the digital storage bucket for the world, with S3, Glacier, Lambda, and EBS in her stack. Before that she ran order-management software at Amazon retail, and she started as an engineer at Microsoft. She remains on the boards of Intel and Confluent, and was previously on Unity's. Guo, a small early Square investor who also worked with Jack Dorsey at Twitter, calls herself a huge admirer of the leadership.

Asked for a defining moment from almost a decade at Square, Henry does not pick a single operational peak. When she joined, the company was still "the little white reader company," the farmers-market payments shop. Cash App was barely a thing, and crypto was not a thought. The business grew from that dongle to merchants of many sizes, including large stadiums and multinationals. The day that still makes her tear up a little is the IPO. Outside the New York Stock Exchange the sign was covered in small-business logos, with a note that the neighborhood was going public. That, she says, sums up a mission-driven company that gave countless small businesses tools and technology previously available to the Amazons and Walmarts of the world.

Guo had watched Dorsey run Square and Twitter in radically different ways. Square was always more GM and business-unit driven; Twitter was often functionally oriented. Henry's own career has switched back and forth for decades: engineer, then engineering management, back to individual-contributor product manager, a general-management role at Microsoft, a functional role at Amazon, a GM seat with a P&L, a functional role again when she joined Square. She prefers the end-to-end, multifunctional GM jobs because they use more parts of the brain. As you get more senior, the work becomes how you both instigate and resolve conflict, keeping the organization on the creativity-and-execution threshold. Those problems, for her, are simply more interesting when they are multifunctional and multidimensional.

Payments companies have used machine learning for years, Guo notes, especially in fraud. Henry splits the story in two. In financial services, ML has been core for a long time. Square has millions of small customers and no one-to-one relationship with most of them, so models have to decide who is a good actor and who is a bad one, whom to target with a product, who will find a marketing offer useful. Risk, fraud, cross-sell, and growth have been internal applications for years. Even before the latest AI shift, Square was using GPT-2 inside Square Messages as a virtual assistant to help merchants answer customer inquiries.

What excites her now is how much better the tools have gotten, and how broadly they can bring expert help to a much larger audience. Talk to small-business owners and most will say they know they should be doing marketing, that reaching customers would drive more business. Then the day ends. They work the shop, go home, take care of family, and at 8 p.m. face a choice: chill for a minute, or spend the next three hours making an image and writing campaign copy. They got into this because they love cupcakes, not because they like writing email marketing. It is too hard, it takes too much time, and they are not experts.

That pattern, multiplied across many chores, is what she calls massive whitespace. If the thing were easy enough, accessible enough, cheap enough, customers would say it would be huge for them. It was not. Hiring a marketing consultant was too expensive, the ROI was not there, so the demand sat underwater. Word processors used to be a product for secretaries; now everyone uses one and the TAM exploded. Square itself is the same curve: how many tiny businesses could accept a credit card when it was too expensive and too hard? Make it better, cheaper, faster, and latent demand appears. Henry's twist on the jobs debate is that some work may be taken, but a lot of it is work that was not getting done at all and now can be: the marketing campaign, a real logo, copy that actually describes the thing.

Guo thinks the effect of generative AI on e-commerce, commerce, and fintech is under-discussed, and she wants the whitespace Henry sees from having adopted GPT-2 so early. Henry's view is that almost every aspect of a small business will find applications. Some of the finance and quantitative pieces are not quite there, because numerical hallucinations have been a bit worse, but that is moving fast. The largest e-commerce players, Amazon and Walmart, have already invested heavily. Online retail had no choice: to operate in e-commerce you must digitize, with images, descriptions, and inventory tracking. In-store and local commerce has no such forcing function, so the rate of digitization significantly lags. The root problem is the old one: digitizing is too hard or too expensive. Generative AI's offer is to make it something like 10x faster and 10x cheaper.

Square heard from businesses that wanted to sell online and did not have product photos. The first version of Photo Studio was heavy: a physical studio in Brooklyn with a 360 camera. Merchants shipped products in, and Square shot professional-grade photos, removing a blocker to getting online. The next version was an iPhone app using less mature image detection and generation to remove backgrounds, so nobody had to ship, and cost came down. With the latest models the barriers have fallen further. The same demand, three solutions, each lighter.

That is the selling side. She sees as much room in the back office: employee management, communication tools, finances, cash-flow prediction, top-selling-product analysis. The data is often already there. Most of these owners are not MBAs. The line in the industry is that they got into business because they like working *in* their business, not *on* it. The interesting part is the craft, the customer, the hospitality. The advances she is watching make it easier to operate with better expertise and less time on the business side of the business.

Gil has been surprised by how much latent demand there is for content generation in every business context. A year earlier, Midjourney and Stable Diffusion were easy to file as cute novelties: how many artists are there, really, who want art for art's sake? Then the same models showed up in product photography, short-form video, spokesperson clips. People who wanted to skip the Brooklyn studio, stay off camera, or shoot attractive marketing video at a hundredth or a thousandth of the cost turned out to be a large set. Many of those demand categories, he notes, do not feel like traditional software businesses. Henry's reply is that this is what the industry is for. Increase ease of use and accessibility and you unlock whitespace, over and over.

Guo wants the GPT-2 origin story, because few companies were wringing usefulness out of messaging models that early. Henry would like to say it was more strategic than it was. Vinod Khosla, then on Square's board, called: a company in his portfolio was being courted for acquisition, would she talk to them. She met two Stanford machine-learning PhD founders, got excited about the technology and the people, and Square acquired the company around 2018. They were put to work stitching together customer-facing experiences. The goal was the same as always: apply technology so merchants get time back for the parts that matter. That team is still going, still moving down the model line.

Aside from AI, what else is changing in e-commerce? Digitization, and even more, integration. Watch a merchant's team work, in person or online, and even the digitized parts are multiple browser windows, cut-and-paste from one tool to another, download-and-email. Manual steps are worse. Henry expects the classic bundling and unbundling cycle. Right now a lot of the landscape is unbundled, best-of-breed pieces. The next phase is bundling, because a bundle addresses integration and can usually be priced below the sum of the parts. Categories that got created in the jump from in-person to e-commerce will consolidate, and the lines between them will blur.

Henry was the first GM for AWS storage and later a much wider computing brief. Does she see AI as a new wave of cloud? It has new aspects. Cloud was historically very CPU-intensive, with some GPU. Training has exploded GPU demand, and that compute is additive, not a replacement for existing workloads. Companies such as Square and Block are still figuring out how to apply the technology and where to train. The race is heating up. In many ways, she says, it is a land grab, with lots of players trying to figure out how to grab land.

Guo notes that this is roughly year one or two of public access to large foundation-model services, and the interface is a single natural-language call, versus AWS's joke-length catalog. Will we get a wealth of services, or just ever more complicated prompts into one model? Probably both, Henry says. SQS was technically AWS's first service; S3 was the one that really landed, with an incredibly simple API of a handful of REST operators. It became not only a first-class product with direct customers but a foundation: trace the call stack of most AWS services and you find S3. OpenAI started relatively simple and is adding capability while also simplifying call patterns, with threads and messages. She expects more core capabilities, more simplification, and some specialization. Amazon Bedrock, a single service that hosts a variety of models, is already a bundle of that kind. We are still super early.

Two of the three major clouds already sit close to a foundation-model company: Google building Gemini, OpenAI aligned with Microsoft. Does Amazon need its own paired model? Henry points out Amazon has the Anthropic investment, Google has also put money into Anthropic, and Microsoft has said it is doing some of its own work alongside OpenAI. Everyone is placing multiple bets: build, buy, partner. A monolithic world looks reasonably unlikely from the customer's side, because an enterprise on your compute may want several models. Cloud marketplaces let customers spend committed AWS (or equivalent) dollars on third-party products and still hit their commit, so the business models intermesh.

Guo remembers asking Henry, in her first year at Square, something Hadoop-related. Henry looked at her and said Amazon loves open source: we make more money off open source than any of the open-source companies do. Guo still files it as one of the most terrifying business conversations of her career, having been investing in those companies. Henry's current view is that enterprises really do want open-source models, because they are not a black box and could theoretically be in-housed, and because developer communities stay passionate. Confluent found a commercial path around Kafka. Hadoop never quite got the commercial piece working, and not enough differentiation from what cloud providers could ship. Clouds have default customer demand: launch a service on the open source and existing customers show up, giving feedback immediately. The cat-and-mouse continues. Both sides will exist.

From the Intel board, Henry offers a wave-by-wave frame. Microcomputers had Intel and AMD, mobile had Arm and Qualcomm, AI has Nvidia on GPUs. Each wave has a clear one and two; there is a three and a four, but in semis it is really hard to be number three and you are dead if you are number four. Nvidia is selling the most GPUs. She does not think the area will be a monopoly, and the number-two seat is still open, with companies running at it and perhaps later challenging number one. As tooling moves up the stack, work that was coded to the metal shifts into tools, which is part of what creates room for a one and a two. Guo adds the Wintel rhyme: GPUs increasingly optimized for transformers, transformer communities optimizing for GPUs, a possible virtuous cycle on top of CUDA and interconnect. The uncertainty is real. AI-chip startups bet that GPUs are great at matrix multiplication and not specifically tuned to transformers, while researchers have started talking about attention's limits and trying other architectures. At the start of 2023 Guo would have said the industry was very committed to transformers; by year's end she felt less confident.

Henry quotes Bezos: in a phase of rapid change, focus on the things you know will not change, because betting on what is not fundamentally true is gambling. Semis have been a scale-up story for years. What she finds interesting now is chiplet designs and advanced packaging, starting to look like a microservices architecture: tune smaller components, rip one out and replace it, rather than change the whole thing. More modular designs should accelerate the innovation cycle. Gil's historical rhyme is a burst of experimentation every cycle, as with L1s invented as Ethereum scalability alternatives that then moved to L2s. About 90 percent of the time, he says, it collapses back to the original thing you keep scaling. Darwin selects for what keeps going.

What is next for her? She does not know; that is what she is working on. Her husband retired two years earlier and has been asking her to come out and play, while she has been working 70 to 80 hours a week. Square was the highest up the stack she had ever worked, and her first time on financial products; Intel and Confluent still scratch the itch for fundamental technology. She is still reading, still watching OpenAI Dev Day, still tinkering. Permanent retirement or a sabbatical: she will see. Gil jokes that he invested in all the good ideas this year and will take the next one off. Then he gets serious. ChatGPT is only a year old, Midjourney and Stable Diffusion a year and change; the last year was everyone waking up. Foundation-model incumbents for the next three or four years are roughly knowable, infrastructure is filling in, and the wide-open space is applications, B2B and consumer. The story of the next two or three years, for him, is incumbents wiring AI into existing workflows and a huge chunk of the services economy converting into AI rather than traditional software. Guo is excited that quality thresholds keep getting crossed faster than expected, from 8-by-8 diffusion images to video that leading researchers recently called a pile of unsolvable problems. Application-layer experiments next year can be more sophisticated, and founders still get to ride the capability curve. No retirement for her, either.