Imagine an AI that doesn’t just understand your business but was born from it – trained on decades of internal documents, workflows, and institutional knowledge rather than generic internet data. This isn’t science fiction; it’s the new frontier in enterprise artificial intelligence, and French startup Mistral is betting big that companies want to build rather than just buy. At Nvidia’s GTC conference this week, Mistral unveiled Forge, a platform that lets enterprises train custom AI models from scratch using their proprietary data. But as the race for enterprise AI dominance heats up, companies face complex choices between customization, security, and platform dependencies that could shape the next decade of business technology.
The Build-Your-Own AI Movement Gains Momentum
Most enterprise AI projects fail not because companies lack technology, but because the models they’re using don’t understand their specific business context. Mistral CEO Arthur Mensch says his company’s laser focus on the enterprise is paying off, with Mistral on track to surpass $1 billion in annual recurring revenue this year. “What Forge does is it lets enterprises and governments customize AI models for their specific needs,” explains Elisa Salamanca, Mistral’s head of product. Unlike approaches that merely fine-tune existing models or use retrieval augmented generation (RAG) to layer proprietary data on top, Forge enables companies to train models from the ground up.
This approach could address limitations of more common methods, particularly for non-English or highly domain-specific data. According to Mistral co-founder Timoth�e Lacroix, “The trade-offs that we make when we build smaller models is that they just cannot be as good on every topic as their larger counterparts, and so the ability to customize them lets us pick what we emphasize and what we drop.” Early adopters include Ericsson, the European Space Agency, and Dutch chipmaker ASML, which led Mistral’s Series C round last September at a �11.7 billion valuation.
The Platform Wars: Microsoft’s Strategic Pivot
Meanwhile, the broader AI landscape is undergoing significant realignment. Microsoft recently reshuffled its AI leadership, shifting responsibilities away from DeepMind co-founder Mustafa Suleyman and appointing former Snap executive Jacob Andreou to run the entire Copilot division. This move comes as Microsoft struggles to gain market share against rivals like Google’s Gemini and market leader ChatGPT. Microsoft’s in-house family of models trails the most advanced offerings from Anthropic, Google, and OpenAI, but the company hopes new releases later this year will close the gap.
More telling is Microsoft’s strategic alliance with Anthropic, where Anthropic’s general-purpose AI agent Cowork will be integrated into Microsoft’s AI assistant Copilot. This marks a d�tente between two companies that were heading toward competitive conflict over AI in enterprise software. Microsoft’s Copilot has had underwhelming adoption with only 15 million paid seats (about 3% of Office users), while Cowork has gained traction as a poster child for AI agents since its debut earlier this year. For Microsoft, this partnership addresses a critical need: “We must achieve true self-sufficiency by building our own powerful models and reducing our reliance on OpenAI,” Suleyman told the Financial Times last month.
The Sovereign AI Imperative and Economic Realities
Beyond corporate strategies, governments are driving another significant trend: sovereign AI. This governmental strategy to secure domestic AI infrastructure – including servers, data centers, and AI models – is gaining momentum in response to geopolitical risks. McKinsey estimates sovereign AI could account for $600 billion in annual spending by 2030, driven by data regulation and reduced dependence on the U.S. Nvidia is already a key beneficiary, with $30 billion in sovereign customer revenue representing 14% of its group total in its last fiscal year.
The economic implications are substantial. If Nvidia captured a quarter of potential physical sovereign spend, its earnings would increase by roughly half at its current 75% gross margin. But this trend comes with duplication costs for countries and raises questions about efficiency versus security. As companies like Mistral position themselves to serve government clients with customizable solutions, they’re tapping into a market where control and security often trump cost considerations.
The Identity Challenge: Proving Humanity in an AI World
As AI agents become more sophisticated, a new problem emerges: how do we ensure these agents represent actual humans? Identity startup World thinks it has a solution with World ID, a cryptographically secure, unique online identity token based on iris-scanning technology. With Agent Kit, World wants to let users tie their confirmed identity to any AI agent, letting it work on their behalf across the internet in a way other parties can trust.
The challenge is adoption. While World claims nearly 18 million unique humans have verified their identities, that’s just a fraction of global internet users. “The trick, of course, is getting a critical mass of people who use AI agents to get their iris scanned for a World ID in the first place,” notes Kyle Orland of Ars Technica. Until then, the chicken-and-egg problem of assigning a unique identity to every online human – and by extension to the agents they might deploy – will remain elusive.
The Innovation Paradox: AI’s Impact on Research
While businesses focus on practical applications, AI’s impact on scientific research reveals a paradox. Research from Tsinghua University shows that while AI adoption helps scientists publish three times as many papers and attract five times more citations, it simultaneously reduces the number of topics studied by 5% and decreases researcher collaboration by 24%. AI amplifies research in data-rich areas but neglects data-poor frontiers, potentially risking long-term scientific progress.
This pattern mirrors what enterprises might experience: AI can optimize existing processes but may not foster truly novel approaches. As companies build custom AI models, they must consider whether they’re creating systems that reinforce existing knowledge patterns or enable breakthrough thinking.
The Road Ahead: Customization vs. Standardization
The enterprise AI landscape is fragmenting into competing approaches. On one side, companies like Mistral advocate for bespoke solutions built on proprietary data. On the other, platform giants like Microsoft and Google push integrated ecosystems. Meanwhile, governments pursue sovereign AI strategies, and identity solutions like World ID attempt to solve emerging trust problems.
For businesses, the choice isn’t simple. Custom models offer better understanding of specific contexts but require significant investment and expertise. Platform solutions provide integration and scale but may not address unique needs. As AI continues to evolve, the most successful enterprises will likely blend multiple approaches, building custom intelligence where it matters most while leveraging platforms for broader capabilities. The question isn’t whether to build or buy AI, but rather what to build, what to buy, and how to ensure it all works together securely and effectively.