News Analysis
White Paper
The Future of Leadership Development: AI-Driven Manager Enablement
Microsoft launched Frontier Company on July 2, a $2.5 billion unit that embeds 6,000 engineers inside customer companies to build out their AI systems.
The platform runs OpenAI, Anthropic, Microsoft AI, open source or specialized industry models "without ceding control to any one of them," wrote Judson Althoff, commercial business CEO, noting that customers shouldn't be locked into a single model any more than a single technology vendor. Client data and IP are never used to train models in ways that would help a competitor, he continued.
Microsoft has not said whether the $2.5 billion is new capital or reallocated from existing consulting budgets, nor has it specified a spending timeline.
Rodrigo Kede Lima, a six-year Microsoft veteran, will lead the unit, according to reports in Technology Magazine. Accenture, Capgemini, EY, KPMG and PwC have been brought in to scale it globally.
Microsoft isn't the only firm adopting this approach. Anthropic officially launched Ode, a $1.5 billion joint venture with Blackstone, Goldman Sachs and Hellman & Friedman, on July 15. OpenAI launched a comparable forward-deployed engineering venture in May, and AWS committed $1 billion to its own initiative two days before Microsoft's announcement.
What Does Free Engineering Actually Cost?
Enterprises increasingly worry that leaning too heavily on frontier AI labs could arm future competitors, analyst Patrick Moorhead told Reuters. Sensitive workflows sent to those labs may hand them insight into industries they could later enter themselves, he said.
Microsoft has already conceded this once. Althoff admitted the company's choice to bind the original Copilot exclusively to OpenAI's models was an error. Customers ultimately care more about the combination of their data and a model than about any particular model, he said.
Free deployment functions as the customer acquisition cost, and Azure consumption meters generate the payback, the same play Microsoft ran with Azure migrations a decade ago, said Lane Shelton, director of advisory services for Directions on Microsoft, an independent research firm with no Microsoft partnership.
"An embedded Frontier engineer co-designing your AI systems isn't just engineering help," Shelton said. It's Microsoft's roadmap arriving as someone else's architecture, in his account, and customers should reckon with that before a contract renews, not after it already has.
Where Does Lock-In Happen?
Whoever writes the reference architecture owns the roadmap, said Vladimir Beskorovainyi, an enterprise AI architect and CTO.
If the model layer is swappable, is the staffing layer where the dependency gets built?
An embedded engineer doesn't need to push Azure outright, Beskorovainyi said. They only need to make the default choice at 50 small forks along the way, from the queue and identity layer to the vector store and eval harness.
Each default is defensible alone. Together, they form a gravity well. "Lock-in stopped being a licensing clause and became an architecture diagram," Beskorovainyi said.
Shelton and Beskorovainyi are describing the same mechanism from different vantage points. Shelton is diagnosing intent: Microsoft's business model is built to convert free engineering into locked-in Azure consumption, the same commercial logic that powered the cloud migration era.
Beskorovainyi is diagnosing mechanics. The lock-in happens regardless of intent, through the accumulated weight of routine technical decisions.
Beskorovainyi recommended asking what it would cost, in engineer-months, to swap the model provider or leave the cloud. If the organization can’t produce that number, the company doesn't have model neutrality, he said.
Open-weight models running on a customer's own hardware already hit roughly 85% accuracy on enterprise text-to-SQL tasks in Beskorovainyi’s benchmarking. That's evidence that a credible second stack exists for most workloads, he said. The question is whether the architecture a Frontier engineer builds preserves the customer's ability to reach it.
Who Can Verify Platform Neutrality?
Model-diverse platforms are only as neutral as a customer's ability to verify them, and few customers actually can.
Classic hyperscalers such as Microsoft and Google, and newer labs such as Anthropic and OpenAI, are moving toward management consulting and systems-integration territory, said Artem Shitov, a knowledge specialist at McKinsey.
Enterprises want new AI infrastructure implemented, and companies that own the technology increasingly want to do that implementation themselves, Shitov said.
"Platform neutrality isn't something a vendor declares," Shitov said. It has to be something a customer verifies, and that's gotten harder as the enterprise AI stack has expanded past model selection into retrieval-augmented generation, agent orchestration, evaluation pipelines and custom hardware for on-premises deployment.
Most enterprises don't have the in-house depth to audit any of that. Neutrality claims, as a result, either get checked by an independent third party or taken on faith.
Engaged, But Not Convinced: What's Eroding Beneath Your Strongest Engagement Numbers
Watch Now
Retention Is a Communication Problem: New Data from 6,000+ Employees
Watch Now
Microsoft isn’t alone in offering multi-model support, Shitov said. Cloud infrastructure players with diverse stacks routinely serve multiple vendors alongside their own models, citing Microsoft's investment in OpenAI alongside its hosting of Anthropic's Claude on Azure and its own MAI and Phi lines, and Google's Gemini running alongside partnerships with both Anthropic and OpenAI, he said.
What Should Enterprises Ask For?
If verification is out of reach, contracts become the fallback. Beskorovainyi recommends asking for three assurances when negotiating the contract to avoid lock-in:
- Get IP and design documentation vesting with the customer as it's produced, not at the end of the engagement.
- Define an exit deliverable on day one, including runnable infrastructure-as-code and model-swap instructions tested before final payment.
- Have a portability service-level agreement, requiring named workloads to demonstrably run on an alternative stack at least once a year. If a vendor resists that clause, that shows whether the system is neutral.