Here’s what’s on our plate today:

• 🧪 When the foundation eats the building.

• 📰 The Marxist AI boss, Eclipse's $2.5B Cerebras win, and Alexa goes shopping.

• 💬 Prompt of the Day: Audit your SaaS moat against OpenAI's latest releases.

• 🗳️ Poll: where does the value land as foundation models move up the stack?

Let’s dive in. No floaties needed…

The Laboratory

TL;DR

• Platform turned competitor: OpenAI’s May 7 voice, translation, and transcription release bundles what previously required multiple vendors into one API, directly competing with its own ecosystem partners.

• The economics are obvious: Voice AI topped $22B, contact centers burn $300B annually, and AI calls cost $0.40 versus $7-$12 for humans, which is too lucrative to ignore.

• Billions of VC at risk: ElevenLabs, Vapi, and Deepgram raised heavily on the assumption that foundation providers would stay in the infrastructure lane, but that assumption is broken.

• Enterprise moats still matter: Compliance, legacy integrations, and industry customization remain gaps that general-purpose models can’t easily close.

• The real stakes: If pricing shifts to outcome-based models as predicted, the line between providers and app companies blurs entirely, and startups without deep enterprise roots risk being absorbed by the platforms they depend on.

When the foundation eats the building

An interesting characteristic of the business world is that it is fundamentally a tightly interconnected network of companies working in sync with one another, each layer depending on the success of the others above and below it.

One of the best examples that comes to mind is the Android operating system. While Google provides the software foundation, smartphone manufacturers build devices around it, developers create applications for it, telecom providers sell connectivity through it, and entire businesses emerge on top of the ecosystem Google created.

Many business relationships, however, become complicated when the company supplying the foundation begins moving up the stack itself. The moment the platform owner starts building products that resemble what its own ecosystem partners are creating, cooperation quietly turns into competition. It’s like Google making an Android device that competes with Samsung’s smartphones. That is increasingly the case in artificial intelligence.

On May 7, 2026, OpenAI released three real-time audio models through its developer API, including GPT-Realtime-2, a voice model with GPT-5-class reasoning capable of holding context, calling tools, and handling interruptions mid-conversation; GPT-Realtime-Translate, which translates speech across more than 70 input languages into 13 output languages in real time; and GPT-Realtime-Whisper, a streaming transcription model.

And while these releases may appear to be simple infrastructure upgrades for developers building voice-powered applications, they also reflect OpenAI’s broader shift from supplying the intelligence layer beneath other companies’ products to offering capabilities that increasingly resemble complete end-user solutions. In effect, OpenAI is moving into the same markets where many of its own enterprise customers already operate.

The pattern, not the product

The release from OpenAI is only the latest step in a broader strategic shift that began earlier this year with the launch of Frontier, the company’s enterprise platform for building, deploying, and managing AI agents. OpenAI positioned Frontier as infrastructure for businesses, capable of connecting to internal systems, executing tasks autonomously, and operating with enterprise-grade security. At the same time, the company has increasingly framed its strategy around supporting a broader ecosystem of partners, even as its own products continue to expand into the same enterprise workflows many of those partners are building their businesses around.

Yet the product itself points toward something more competitive. Frontier is built around the idea of AI agents functioning as digital coworkers embedded directly into day-to-day business operations, handling workflows that traditional enterprise software platforms were designed to support. In doing so, OpenAI moves far closer to the same workplace software market long dominated by SaaS (software-as-a-service) companies, blurring the line between infrastructure provider and application-layer competitor.

Where the money actually sits

The shift makes clear economic sense for OpenAI because the voice AI market alone is on pace to surpass $22B in 2026, while companies worldwide still spend roughly $300B annually on contact center operations, making the sector one of the largest automation opportunities in enterprise technology. AI voice agents are especially attractive because they dramatically reduce costs: an automated customer service call costs around $0.40, compared to $7-$12 for a human agent.

And for foundation model providers, the real value increasingly lies not just in supplying the underlying AI models, but in owning the applications, workflows, and business processes where those savings are actually created. OpenAI’s May 7 audio release clearly illustrates that shift, aiming to provide capabilities that would previously have required multiple vendors across speech recognition, translation, transcription, reasoning, and workflow orchestration.

However, while the strategy makes clear business sense and could prove highly profitable for OpenAI, it also places growing pressure on companies already operating in the space, where billions of dollars in venture capital and market value are at stake.

Billions already in the crosshairs

The companies currently occupying the voice AI application layer have attracted enormous amounts of capital in a very short time. ElevenLabs reached an $11B valuation earlier this year. At the same time, companies such as Deepgram, Parloa, PolyAI, Retell AI, and Vapi have all raised heavily amid surging demand for AI-powered voice automation. According to CB Insights, voice AI startups raised more than $2.1B globally in 2024, with hundreds of millions more flowing into the sector in early 2025 alone.

Much of that growth rested on a key assumption: that foundation model providers would remain infrastructure companies, leaving the application layer to specialized startups. Companies like Vapi, for example, built orchestration platforms that combined separate speech recognition, language, and voice-generation systems into a single service. That model becomes harder to sustain when OpenAI begins bundling many of those same capabilities into a single API endpoint at far lower cost.

Beyond startups, the pressure is increasingly spreading across the broader SaaS industry. As OpenAI and Anthropic pushed deeper into the application layer, investors began questioning the economics of traditional software businesses built around per-seat pricing. The concern was straightforward: if AI agents can perform the work directly, companies may no longer need to pay for as many human-operated software seats.

What the models still cannot do

There is, however, a credible counterargument to the idea that foundation model providers will inevitably swallow the application layer. Goldman Sachs CEO David Solomon described the February selloff in AI-related stocks as “too broad,” arguing that the market was failing to distinguish between companies likely to be disrupted by AI and those positioned to benefit from it.

Speaking at a UBS conference, Solomon said there would inevitably be “winners and losers,” but added that many companies would adapt and continue to perform well. Dan Ives similarly described the selloff as a “generational opportunity” for investors, while JPMorgan argued that market sentiment had become excessively pessimistic.

The reason for that optimism lies in execution. OpenAI’s Realtime API is still fundamentally a model endpoint rather than a fully enterprise-ready service, which does not automatically provide the compliance, legacy integrations, or industry-specific customization required in heavily regulated sectors.

Companies like PolyAI have built businesses around those gaps, using specialized customer-service data and enterprise integrations to create systems that reportedly resolve more than 80% of calls without human intervention. That kind of operational expertise is difficult to replicate with a general-purpose AI model alone.

Traditional SaaS companies argue that their value goes far beyond the AI models themselves. Many believe their strengths lie in governance, cybersecurity, data ownership, compliance, and enterprise access controls, all of which become more important as AI systems spread across large organizations. They also argue that most enterprise workflows are unlikely to become fully autonomous anytime soon, meaning businesses will continue operating in hybrid environments where humans and AI systems work together.

Some analysts believe markets may also be overstating the threat of foundation model builders moving in and destroying the companies operating in the application layer.

The Bank of America Global Research described the market’s two biggest fears as “internally inconsistent”: on one hand, investors worry AI spending will become economically unsustainable, while on the other, they fear AI will become so effective that it destroys the value of existing software companies entirely.

When pricing becomes the product

However, a more likely outcome may be the one IDC predicts: that by 2028, roughly 70% of software vendors will move away from traditional per-seat pricing models and instead charge based on measurable business outcomes, usage, or organizational capabilities.

Gartner reports that 40% of enterprise SaaS contracts already include some form of outcome-based pricing, up from just 15% two years ago. If that transition accelerates, the distinction between model providers and application-layer companies may become less important than many investors currently assume.

From platform partner to platform rival

And OpenAI’s own language increasingly reflects that shift. At the launch of Frontier, the company described AI agents as evolving into “true AI coworkers,” a framing that points toward a future where the real competition is no longer simply about building the best model or the best software interface, but about controlling the workflows and business outcomes those systems produce.

In many ways, the trajectory resembles what happened when Google, after years of supplying Android as the software foundation for the smartphone industry, eventually began building its own Pixel devices. The move did not destroy Android’s ecosystem. Still, it fundamentally changed the relationship between Google and the companies that depended on it as partners became competitors, even while continuing to rely on the same underlying platform.

The same dynamic is now emerging in AI, where, as foundation model providers move further into the application layer, the question is no longer whether they will compete with the companies built on top of them, because they already are. The larger question is which parts of the AI stack will continue to hold independent value once the foundation itself becomes capable of absorbing more of the product layer above it.

Tuesday Poll

Prompt Of The Day

Bite-Sized Brains

• The Marxist AI boss: A new chatbot is being marketed as a Marxist alternative to corporate AI assistants, a quirky reminder that political branding is now arriving in the AI productivity layer.

• Eclipse's $2.5B Cerebras win: Robotics startup Eclipse landed a $2.5B Cerebras deal as it pushes deeper into its physical-world AI thesis, signaling fresh momentum for hardware-first AI bets.

• Amazon's Alexa goes shopping: Amazon is fusing Alexa with Rufus to turn its voice assistant into a full-blown AI shopping agent, putting agentic commerce front and center.

The Toolkit

• Deepgram: Speech-to-text API built for scale, handling real-time transcription and voice intelligence for production apps.

• Descript: AI-powered audio and video editor that lets you edit recordings by editing the transcript like a doc.

• Drift.ai: AI-powered conversational marketing platform that turns website visitors into qualified pipeline through automated chat. 

Rate This Edition