Here’s what’s on our plate today:

• 🧪 OpenAI tested everything. Only three business models actually stuck.

• 🗞️ SAP's HR agents, Mythos on notice, and Amazon enters drug discovery.

• 📊 Friday Poll on whether OpenAI's superapp bet lands in time.

• 🛠️ A weekend to-do testing the products at the center of the story.

Let’s dive in. No floaties needed…

The Laboratory

TL;DR

• Ads worked, everything else is complicated: The ad pilot crossed $100M in annualized revenue within six weeks, with 600+ advertisers paying ~$60 CPM. Turns out 900M free users are worth something if you monetize their attention.

• Sora and Instant Checkout died for different reasons: Sora collapsed because consumer pricing cannot subsidize GPU-intensive video. Instant Checkout failed because real-time catalog management requires infrastructure that OpenAI tried to skip.

• The ‘side quests’ memo is the real story: Fidji Simo told employees to stop chasing ancillary products, then announced Atlas, ChatGPT, and Codex would merge into a single superapp. The Edison phase is over.

• Only three AI business models actually scale: Ads, subscriptions, and enterprise APIs. Commerce, consumer video, and hardware keep hitting infrastructure deficits and unsustainable unit economics.

• The IPO math demands focus, not breadth: With a potential $1T+ valuation and ~$9B in annual costs, OpenAI cannot keep subsidizing experiments. If consolidation works, it becomes a productivity platform. If not, the growth narrative looks like the filaments Edison discarded.

Why OpenAI’s do-everything strategy may not be enough

A famous quote often attributed to Thomas Edison is that, “I have not failed. I’ve just found 10k ways that won’t work.” Behind the line lies the reality of his work on the light bulb, where he and his team tested thousands of materials for the filament, most of which failed. Rather than treating these as setbacks, Edison saw each attempt as progress, a way of ruling out what didn’t work. His eventual success with a durable carbon filament was not a single breakthrough, but the result of systematically eliminating thousands of dead ends.

Today, a company with a multi-billion-dollar valuation appears to be taking Edison’s approach of testing thousands of materials to find the one that best suits its needs.

Over the past few months, OpenAI has been operating less like a focused AI research lab and more like a conglomerate running a dozen startups simultaneously, launching products across advertising, e-commerce, video generation, web browsing, hardware, and adult content. Some of these bets are paying off: the advertising pilot crossed $100M in annualized revenue in under two months, and the company hit $25B in annualized revenue by February 2026. Others have already been abandoned.

Among the projects it shut down were the Sora video generation app in March 2026, just six months after launch. It walked back its native e-commerce checkout after struggling to onboard merchants and maintain accurate product data. The planned ‘adult mode’ for ChatGPT has also been delayed twice. And hardware, designed with Jony Ive, is still in prototyping with delivery timelines slipping toward 2027.

The pattern reveals a company testing the outer limits of what an AI platform can become while simultaneously preparing for a potential IPO at a valuation that could exceed $1T. However, the tension between experimentation and financial discipline has now produced a visible correction: in March 2026, the CEO of Applications, Fidji Simo, told employees to stop pursuing “side quests” and refocus on coding tools and enterprise users.

Three days after that internal address, OpenAI confirmed it would merge Atlas, ChatGPT, and Codex into a single desktop superapp. The message was unmistakable: the Edison phase is ending. The company believes it has found its filaments, and everything else is being switched off.

The experiments that worked

Not everything OpenAI tried was a misfire, and the advertising business in particular has been a quiet vindication of the company’s scale.

When OpenAI announced in January 2026 that it would begin testing ads for free and Go-tier users, the response from the AI community ranged from resignation to alarm. Here was the company that had once styled itself as a nonprofit research lab, now selling ad impressions against the conversations of 900M weekly active users. But the commercial logic was hard to argue with. The overwhelming majority of those users, roughly 95%, had never paid OpenAI a cent. Against an estimated $9B in annual operating costs, the math was unsustainable without a second revenue stream.

The ads went live on 9 February 2026, and within six weeks, the business surpassed $100M in annualized revenue with more than 600 advertisers signing on. ChatGPT ads launched at roughly $60 per thousand impressions, a significant premium over traditional display advertising, reflecting the platform’s unusual intimacy: users are not scrolling a feed but conducting a conversation, and an ad surfaced in that context carries a different weight. WPP has estimated that OpenAI could pull in between $500M and $800M in its first year of advertising.

Self-serve advertising tools that remove the current $200k minimum commitment are expected to launch in April 2026, potentially opening the floodgates for small and mid-sized businesses. If that happens, advertising could move on from being a supplement to subscriptions to being a co-equal pillar of OpenAI’s business model.

Codex, meanwhile, has emerged as the company’s clearest enterprise product. By January 2026, OpenAI’s coding tools were generating just over $1B in annualized revenue, making the case that developers, not consumers browsing TikTok-style video feeds, represent the highest-value audience for AI. The decision to fold Codex into the superapp is less a pivot than an acceleration of what was already working.

The experiments that didn’t

But in OpenAI’s playbook for every advertising pilot that clicks, there is a Sora.

When Sora 2 launched as a standalone app in September 2025, complete with a TikTok-style vertical feed, it briefly topped the App Store charts. To ensure Sora's success, OpenAI secured a high-profile partnership with Disney, signed just months before the shutdown, and the product represented a genuinely ambitious attempt to build a consumer creative platform powered by generative video.

However, the platform’s collapse was swift and total. From a peak of 3.3M downloads in November 2025, usage plummeted to 1.1M by February 2026, generating a mere $2.1M in lifetime revenue. The technology did not fail; Sora could produce striking video, but despite the technical prowess, the economics did not make sense.

Unlike text generation, video generation consumes compute at a rate that OpenAI’s consumer pricing cannot absorb, and the company ultimately concluded that those GPU cycles were better spent elsewhere. When OpenAI announced the shutdown on 24 March 2026, it indicated the research team would pivot toward “world simulation” for robotics rather than consumer video. The creative app was dead, but the underlying research would live on in a more commercially defensible form.

Apart from Sora, OpenAI also pulled the plug on its e-commerce experiment. And even though Instant Checkout was a different kind of failure, rooted not in compute costs but in infrastructure, it reflects the company’s push to ensure it can loop in investors when it goes public.

Launched in September 2025 as an in-chat purchase feature built with Stripe, the e-commerce venture promised to let users buy products without leaving the conversation. The vision was seductive: ask ChatGPT for a recommendation and purchase it in the same breath.

In practice, the reality was far messier. Only about a dozen Shopify merchants out of millions on the platform have integrated with Instant Checkout. For OpenAI, the fundamental challenge was data: while it could scrape retailer websites, this meant stock status, delivery timing, and shipping costs were often inaccurate or outdated. All of which it could not manage or invest in building. Besides, both e-commerce and video generation presented regulatory challenges that would consume valuable time and energy that can be invested in more profitable ventures.

Where the model breaks: infrastructure, not intelligence

There was also a regulatory gap that the company had not closed. As of February 2026, OpenAI had not built a system for collecting and remitting state sales taxes, which represented a live compliance problem for any transaction routed through ChatGPT’s checkout. By early March, OpenAI confirmed it was stepping back from native checkout entirely, shifting purchases to retailer apps that merely connect to the chatbot. The commerce ambition wasn’t abandoned so much as deflated into something far more modest.

For outsiders, OpenAI’s strategy can be viewed in two ways. On one hand, it reflects a deliberate strategy: in a fast-moving field, it is better to try and fail than to miss the next big opportunity, with each product narrowing the path toward what works. On the other hand, it points to mounting financial pressure, where rapid launches help sustain a growth narrative needed to justify massive funding and delayed profitability. The reality likely sits in between, with the company now shifting toward discipline, consolidating products, and focusing on a clearer identity as a productivity platform.

The limits of AI as a business model

On a broader level, OpenAI’s current strategy reveals a hard truth about the AI sector. At the same time, models can do remarkable things; only a few business models—ads, subscriptions, and enterprise APIs—consistently scale. Efforts to expand into areas like commerce, video, or hardware run into real-world constraints, such as infrastructure, regulation, and user behavior. Competitors like Anthropic have leaned into enterprise tools and are growing rapidly, while giants like Google and Apple still dominate key ecosystems. The lesson, then, is that AI capability alone is not enough; turning it into a sustainable business remains the harder challenge.

Just as Edison’s success did not come from random trial and error, but from disciplined experimentation, OpenAI needs to spend time having a focused strategy that could mean the difference between trying everything and building something.

The company and the industry as a whole need to realize that the technology is powerful, but capability alone does not make a business. Success depends on infrastructure, distribution, regulation, and sustainable economics. The companies that win will not be those that try everything, but those that figure out what actually works and build around it. The question now is whether OpenAI can make that shift in time.

Headlines You Actually Need

• SAP's HR agents: SAP embedded agentic AI across recruiting, payroll, and talent management to cut enterprise admin bloat.

• Mythos puts the world on notice: Anthropic's Claude Mythos found a 27-year-old OS flaw in days, and academics are already asking what happens when that spreads.

• Amazon enters drug discovery: AWS launched Amazon Bio Discovery, cutting drug candidate generation from 18 months to weeks.

Friday Poll

Weekend To-Do

• Test Codex on a real task: OpenAI's coding tool is the one product quietly generating $1B+ in ARR. Give it an actual problem from your workflow and see why the enterprise is paying for it.

• Compare Claude vs ChatGPT side by side: Run the same prompt through both. The gap in enterprise perception is real, and you'll feel it.

• OpenAI superapp: Search for coverage of OpenAI's internal consolidation and map it against your own product or business. The same discipline applies.

Meme Of The Day

The Toolkit

• Pika: Fast, social-first AI video tool built for creators who want quick scenes, effects, and remixable clips without pro editing skills.

• Runway: Higher-end AI video platform for more controlled, cinematic generation and real production workflows.  

• CapCut: Consumer-friendly editing stack with AI video tools, captions, speech features, and templates already built for short-form distribution.

Rate This Edition