Direct answer: AI scaling laws — more computing, higher quality data, better infrastructure — have not yet reached their limit, according to Kevin Scott, Microsoft CTO. For companies, the lesson is not training models, but putting that capacity to work in a concrete process: serving, qualifying, and scheduling.
Update (Jun/2026): Kevin Scott's interview on the Training Data podcast by Sequoia Capital is from 2024, but the practical takeaway remains valid — available AI capacity continues to exceed what companies are using.
While the debate continues on how far artificial intelligence will scale, most companies still pay a team to do what AI could already solve alone — respond instantly, qualify a lead, schedule. The AI scaling laws help explain why this frontier moves so fast: the engine behind it is computing at scale. In a 2024 interview for the Training Data podcast by Sequoia Capital, Kevin Scott, CTO of Microsoft, explains where there is still room to grow — and the part that matters for your business is not his infrastructure, but what you do with it. See in 3 minutes which process your company should automate first.
Who is Kevin Scott and Why This Scaling Vision Matters
Kevin Scott is the CTO (Chief Technology Officer) of Microsoft and has led the company's AI strategy for years — the same one supporting the partnership with OpenAI and the cloud infrastructure behind much of the generative AI used in the market today. He describes himself as "short-term pessimist, long-term optimist": skeptical of immediate hype, convinced the capacity curve is far from capped.
The public point anchoring this text is straightforward: for Scott, the industry has not yet reached diminishing marginal returns at scale. In other words, continuing to increase computing power and quality data still produces better models — it’s not a dead end. This is the verifiable source; the rest of this article translates it into your operation.
In field practice: the most common error we see among managers is waiting for the "next generation" of models to start. Available capacity, according to Scott himself, is more powerful than how companies are using it. The bottleneck is rarely the model — it’s the lack of a clear process for it to solve.
The three levers of scale: computing, data, and infrastructure
When Scott talks about scale, he’s not talking about just one thing. There are three levers that reinforce each other, and understanding each avoids buying the wrong argument:
- Computing — the processing power (GPUs, data centers) that enables training and running larger models. It’s the most expensive lever and the one most seen in headlines.
- Data — here is the key turning point in his view: quality matters more than quantity. A well-curated "curriculum" of data produces smarter models than simply dumping more text.
- Infrastructure — the layer that serves models reliably and cheaply enough for real-world, everyday production use.
Scott also notes an important shift: as frontier models mature, the effort moves from training (creating the model) to inference (using the model millions of times in concrete cases). For company leaders, this phrase is priceless: the value increasingly lies in applied use, not in who trains the largest model.
What we've learned in operation: the "data quality" lever has a direct parallel in day-to-day work. An AI agent connected to your CRM, your service history, and your calendar responds much better than a generic model — not because it’s larger, but because it has the right context. Raw scale without your business context delivers less than it seems.
Why "scaling laws" don’t last forever — and what comes next
Scott is optimistic but not naive. He reminds us every scaling law has a horizon. The Moore’s Law (the number of transistors on a chip doubling approximately every two years) supported decades of progress but is not eternal; the Dennard scaling, related to energy use per transistor, stopped holding true long ago. Technology history consists of curves that rise, saturate, and are replaced by new architectures.
The practical takeaway is liberating: you don’t need to bet on the next research frontier. You can reap gains from what already exists now. The companies advancing fastest in AI adoption, according to Scott’s observation, are precisely those applying current capacity with judgment and purpose — not those waiting for the next leap.
In field practice: we translate this into a simple rule. Start with the most repetitive and measurable process — in nearly every business, that’s service and qualification on WhatsApp. It’s where current technology already delivers clear return without relying on any future promise. This is the logic behind our process automation: apply what already works, at the point of greatest friction.
From global scale to your WhatsApp: what this means for your company
Here Kevin Scott’s story meets the ground of your operation. The scale he describes — computing, data, infrastructure — is exactly what allows a Digital Employee to run 24/7, handle dozens of conversations simultaneously, and maintain quality. You don’t operate the data centers; you reap their results in a concrete process.
At XMACNA, this capacity has a name and function: the Digital Employee is an AI agent that not only chats but executes an end-to-end process — serves, qualifies, schedules, and records — integrated with the systems you already use. There are already +600 Digital Employees in operation, handling +30.000 messages per day. The results show where the task is repetitive and timing matters: at Rede Supera, the Digital Employee doubled scheduled visits (+100% versus the network’s own control group) and generated +100% effective contacts. At Instituto Mix, the contact-to-visit scheduling rate jumped from 1 in 10 to 6 in 10. These are real data, auditable on the Intelligent Dashboard.
This is the point Scott repeats in another key: models are already more powerful than the usage made of them. The competitive advantage is not in having the biggest model — it is in applying what exists in a process that returns hours and closes more deals. That is why it makes sense to start with WhatsApp support24/7.
In summary
- For Kevin Scott, Microsoft's CTO, AI scale has not yet reached its limit — quality computing and data still produce better models.
- There are three levers: computing, data (quality > quantity), and infrastructure; the value is shifting from training to using (inference).
- Laws of scale saturate and are replaced — so the practical gain comes from applying current capacity, not waiting for the next leap.
- Applied to business, this is the Digital Employee from XMACNA: it serves, qualifies, and resolves on your WhatsApp, with auditable results.
Frequently asked questions
Who is Kevin Scott and what is his position at Microsoft?
Kevin Scott is Microsoft's CTO (Chief Technology Officer) and leads the company's artificial intelligence strategy, including cloud infrastructure and partnership with OpenAI. The scale vision cited here comes from his public interview on the Training Data podcast by Sequoia Capital, in 2024.
What does Kevin Scott say about AI's laws of scale?
That the industry has not yet reached diminishing marginal returns on scale — meaning that increasing quality computing and data still produces better models. He describes himself as "short-term pessimist, long-term optimist" and believes data quality matters more than quantity.
Will the laws of scale last forever?
No. Scott reminds us that every law of scale has a horizon: Moore's Law held for decades, but Dennard scaling, linked to energy consumption, has already stopped. The industry tends to migrate to new architectures. For companies, the lesson is to harvest the gain from capacity that already exists today.
What do AI's laws of scale change for my company?
You do not need to train models or operate data centers. Global scale is what enables a Digital Employee to run 24/7 with quality. The gain comes from applying this capacity to a concrete process — usually support and qualification on WhatsApp.
Where do I start using AI in my company?
With the process that has the most friction and is most repetitive, which is usually support and qualification. XMACNA’s free assessment shows, in 3 minutes, which process to automate first, with no commitment.