This is Scale & Strategy, the newsletter that's like having a Bloomberg terminal, but with personality.
Here’s what we’ve got for you today:
The next AI winner may have the best guardrails, not the best model
The AI industry just remembered costs exist
The next AI winner may have the best guardrails, not the best model
The AI industry loves talking about smarter models.
Most enterprises are worried about something far less exciting: whether they can actually deploy them without creating a security nightmare.
That was the underlying theme of last week's announcements from Microsoft and Snowflake.
Both companies unveiled a wave of products designed to solve two of the biggest bottlenecks holding back enterprise AI adoption: trust and simplicity.
Neither problem is particularly glamorous. Both are enormously important.
The first challenge is trust.
Companies are increasingly comfortable experimenting with AI. They're far less comfortable giving autonomous agents access to internal systems, sensitive data, customer records, financial information, and operational workflows.
That's where both companies focused a significant amount of attention.
Snowflake introduced Horizon Catalog, a governance layer designed to give agents auditable identities, continuous security monitoring, and shared organizational context through a feature called Horizon Context. The goal is straightforward: ensure agents operate with the same understanding of the business as the humans using them.
Microsoft took a similar approach from a different angle. The company launched Agent Control Specification, an open standard for governing agent behavior, alongside Project MDASH, which uses AI agents to proactively identify vulnerabilities and security weaknesses.
The message from both companies is clear: enterprises won't deploy agents at scale unless they understand what those agents are doing, what data they can access, and how they're being controlled.
Trust isn't a feature. It's infrastructure.
The second challenge is simplicity.
A surprising amount of enterprise AI still requires users to understand prompts, integrations, permissions, workflows, connectors, APIs, and enough acronyms to make a compliance officer cry.
That's fine for developers.
It's a problem for everyone else.
Microsoft's answer is Scout, a personal work agent that can proactively handle tasks across tools like Teams and Outlook.
Snowflake expanded its own agent offering, Snowflake Cowork, with Cortex Sense for unified context, User Memory to learn recurring behavior, and Skills that allow users to package and share workflow automations across teams.
The common thread is obvious.
The companies aren't trying to teach every employee how AI works. They're trying to make AI disappear into the workflow.
That's probably the correct strategy.
One reason AI adoption has looked relatively smooth so far is that developers were among the first major users. Developers naturally tolerate complexity because dealing with complexity is part of the job.
The average knowledge worker has a very different threshold.
If agents require extensive setup, constant oversight, or technical expertise, adoption slows dramatically. If they quietly work in the background and solve problems, adoption accelerates.
The broader takeaway from these launches is that enterprise AI is entering a new phase.
For the past few years, most of the industry's attention has been focused on model performance. Bigger context windows. Better reasoning. Higher benchmark scores.
Those improvements still matter.
But enterprises increasingly care about a different set of questions.
Can I trust this system?
Can I deploy it safely?
Can employees use it without specialized training?
Can it integrate with the systems we already have?
The companies that solve those problems may end up capturing more value than the companies that simply build the next slightly smarter model.
That's why Microsoft and Snowflake's strategy makes sense.
They're positioning themselves one layer above the models, building the infrastructure that allows enterprises to actually use them.
In every gold rush, most people focus on the miners.
The larger businesses often get built by the companies selling the equipment everyone else depends on.
Framer helps teams design, build, and launch their marketing sites lightning fast. With the ability to publish hundreds of CMS pages in a single click, operate at a global scale with seamless localization, and even host unified content across multiple domains, teams have never been able to ship faster. Trusted by companies like Miro, Bilt, and Perplexity
Speed without chaos: ship pages and updates faster without turning the site into a fragile set of one-off hacks
Reduce dependency: shift routine brand and marketing work out of product engineering queues.
Production-grade foundation: Run real marketing systems (CMS, SEO, performance optimization) with governance and collaboration
For the past two years, the AI playbook has been simple: use more tokens, run bigger models, spin up more agents, and worry about the bill later.
The bill has arrived.
One theme quietly connected many of the announcements from Microsoft Build and Snowflake Summit this week: efficiency.
Not better benchmarks.
Not larger context windows.
Not more agents.
Efficiency.
That's a notable shift because enterprise AI spending has been operating under a fairly dangerous assumption: if AI creates enough value eventually, the cost structure today doesn't matter.
That logic works for a while. Then finance starts asking questions.
As AI deployments scale across organizations, token costs, inference costs, infrastructure costs, and agent execution costs have started becoming real line items rather than experimental budgets.
Companies are discovering that a workflow that looks magical in a demo can become very expensive when thousands of employees use it every day.
Rob Ferguson, VP of Technology and Strategy at Fireworks AI, put it bluntly this week.
Many companies became obsessed with maximizing token consumption without paying enough attention to outcomes.
The industry even developed its own version of performance theater. More agents. More reasoning steps. More tool calls. More tokens. More everything.
The assumption was that bigger automatically meant better.
Reality tends to be less generous.
A surprising number of enterprise tasks don't require a frontier model burning through enormous amounts of compute. Many don't require an agent at all.
That's why several of this week's product launches focused heavily on cost efficiency.
Snowflake's new Cortex Training platform emphasizes cheaper and faster customization of open-weight models. Adaptive Compute automatically allocates resources in real time to improve infrastructure efficiency.
Microsoft's new MAI-Thinking-1 reasoning model takes a similar approach. At 35 billion parameters, it's dramatically smaller than many of the trillion-parameter systems dominating headlines today.
That wasn't an accident.
Microsoft is clearly betting that many enterprise workloads care more about cost-adjusted performance than absolute benchmark supremacy.
The company pushed that philosophy further on the hardware side with new local AI devices designed to run models directly on user machines.
The idea is straightforward: if a task can be handled locally, why pay cloud inference costs every time?
Microsoft calls that vision "unmetered intelligence."
That's a clever phrase, but the underlying economics are what matter.
Every prompt costs money.
Every agent action costs money.
Every tool call costs money.
Every reasoning step costs money.
At small scale, nobody notices.
At enterprise scale, somebody always notices.
The next evolution of enterprise AI adoption may have less to do with model capabilities and more to do with discipline.
Raj Ramanujam of Dynatrace made what might be the most important point of the week: before asking which AI system to deploy, companies should ask whether they need AI at all.
That sounds obvious.
It's surprisingly rare.
The AI boom created enormous pressure on organizations to prove they were "AI-first." In many cases, that led teams to deploy AI because they could, not because they had clearly defined problems worth solving.
Those experiments generated plenty of activity.
Generating ROI has been harder.
That's why the industry feels different today than it did a year ago.
The conversation is slowly shifting from capability to economics.
Can the model do this?
Can the agent do that?
Those questions still matter.
A more important question is emerging: does the value created justify the compute bill?
The companies that answer that question successfully will likely outperform the companies that simply consume the most tokens.
Because in the real world, nobody gets promoted for generating the largest cloud invoice.
Framer helps teams design, build, and launch their marketing sites lightning fast. With the ability to publish hundreds of CMS pages in a single click, operate at a global scale with seamless localization, and even host unified content across multiple domains, teams have never been able to ship faster. Trusted by companies like Miro, Bilt, and Perplexity
Speed without chaos: ship pages and updates faster without turning the site into a fragile set of one-off hacks
Reduce dependency: shift routine brand and marketing work out of product engineering queues.
Production-grade foundation: Run real marketing systems (CMS, SEO, performance optimization) with governance and collaboration
