In the summer of 2017, a 20-year-old intern at Google Brain in Toronto contributed to a research paper that would fundamentally reshape artificial intelligence. The paper, titled "Attention Is All You Need," proposed a novel architecture called the transformer—a mechanism that would become the foundation for GPT, BERT, Claude, and virtually every major AI model that followed.
The intern's name was Aidan Gomez. He was the youngest of eight authors, working alongside legendary researchers like Ashish Vaswani, Noam Shazeer, and Niki Parmar. While completing his undergraduate degree in computer science and mathematics at the University of Toronto, Gomez had stumbled into one of the most consequential research projects in computing history.
"I was just trying to understand how attention mechanisms worked," Gomez would later recall. "We didn't know it would become the architecture for everything." The transformer's elegance lay in its simplicity: instead of processing sequential data step-by-step like previous models, it could analyze entire sequences simultaneously through attention mechanisms, dramatically improving both speed and accuracy.
Seven years later, that research paper has been cited over 120,000 times. And Gomez, now 28, has built Cohere—an enterprise AI company valued at $6.8 billion that raised $500 million in August 2025 and is preparing for an IPO that could value the company above $10 billion. In an AI landscape dominated by OpenAI's consumer-facing ChatGPT and Anthropic's safety-first Claude, Gomez has carved out a third path: enterprise-focused, privacy-conscious, and optimized for retrieval-augmented generation (RAG) that allows companies to deploy AI without sending sensitive data to external servers.
This is the story of how a math teacher's son from Brighton, Ontario, went from Google Brain intern to leading one of AI's most strategic challengers—and why his bet on enterprise AI might prove more durable than the consumer chatbot wars capturing headlines.
Aidan Gomez grew up in Brighton, a small town on the north shore of Lake Ontario, population 12,000. His father taught math and physics at the local high school; his mother was a librarian, dancer, and artist. The household blended analytical rigor with creative expression—a combination that would later inform Gomez's approach to AI product development.
"My dad would bring home math problems from his classes, and we'd work through them at dinner," Gomez said in a 2023 interview. "My mom taught me to think about systems aesthetically, not just functionally. Good design is elegant design."
Gomez's dual British-Canadian citizenship (his mother was British) gave him access to broader academic networks. He excelled in mathematics competitions and computer science olympiads, winning regional championships that caught the attention of University of Toronto recruiters. In 2013, he enrolled in the university's computer science and mathematics program, studying under Roger Grosse, who would become a key mentor.
By his sophomore year, Gomez was already publishing research papers on neural architecture search and optimization algorithms. His undergraduate work on efficient training methods attracted attention from Google Brain's Toronto office, leading to the internship that would change his trajectory.
When Gomez joined Google Brain in early 2017, the team was wrestling with a fundamental problem in natural language processing: how to enable models to understand context across long sequences of text. Recurrent neural networks (RNNs) and Long Short-Term Memory (LSTM) models processed text sequentially, which created bottlenecks and made it difficult to capture relationships between distant words.
The Google Brain team—led by Ashish Vaswani and including researchers from Google Research—proposed a radical solution: eliminate recurrence entirely. Instead, use attention mechanisms that could directly model relationships between all words in a sequence simultaneously. This would enable massive parallelization during training, dramatically reducing the time and compute required to build powerful language models.
Gomez's contribution focused on the model's architecture design and experimental validation. "I ran hundreds of experiments testing different configurations," he explained in a 2024 podcast interview. "We were trying to figure out: how many attention heads? What dimensions? How deep should the network be? It was tedious work, but each experiment taught us something about how attention scaled."
The paper was submitted to the Neural Information Processing Systems (NeurIPS) conference in June 2017 and accepted in September. By December, when it was officially published, the AI community recognized they were witnessing a paradigm shift. The transformer's efficiency and scalability made it the obvious choice for future foundation models.
OpenAI's GPT (Generative Pre-trained Transformer), released in 2018, was built directly on the transformer architecture. So was BERT from Google. And Claude from Anthropic. And LLaMA from Meta. The eight authors of "Attention Is All You Need" had collectively enabled the modern AI revolution.
For Gomez, then 20 years old and still an undergraduate, the question became: what next?
After graduating from University of Toronto in 2018, Gomez enrolled in a PhD program at Oxford University, studying under Yarin Gal and Yee Whye Teh—two prominent researchers in probabilistic machine learning and Bayesian deep learning. The plan was straightforward: complete the doctorate, continue publishing research, perhaps return to an industry lab like Google Brain or DeepMind.
But by mid-2019, Gomez had become increasingly convinced that the transformer's real value wouldn't be captured in academic papers—it would be unlocked by companies building products for enterprises. OpenAI's GPT-2, released in February 2019, demonstrated the commercial potential of large language models. Yet OpenAI was focused on consumer applications and AGI research. Google was constrained by corporate bureaucracy and conflicting priorities. There was an opening for a company focused exclusively on helping businesses deploy transformers safely and effectively.
"I looked at the enterprise software landscape and realized no one was building foundation models optimized for business use cases," Gomez told investors in Cohere's Series A pitch deck, obtained by The Information. "Companies needed private deployment, industry-specific fine-tuning, and seamless integration with existing data infrastructure. That was a different product than consumer chatbots."
In September 2019, Gomez made the decisive move: he left Oxford (his PhD would eventually be completed in absentia and awarded in May 2024), returned to Toronto, and co-founded Cohere with Nick Frosst (another Google Brain researcher) and Ivan Zhang (an engineering lead from Tensorflow). The name "Cohere" was chosen deliberately—it meant bringing together disparate elements into a unified whole, which captured both the technical function of attention mechanisms and the company's mission to integrate AI into enterprise workflows.
Radical Ventures, a Toronto-based AI-focused VC firm co-founded by Geoffrey Hinton (the "godfather of deep learning"), led Cohere's $40 million Series A round in November 2020. Hinton's involvement provided instant credibility: here was the most respected figure in AI endorsing a 23-year-old CEO's vision for enterprise transformers.
From the beginning, Cohere's product strategy diverged sharply from OpenAI's and Anthropic's approaches. While those companies focused on building the most capable general-purpose models and monetizing through API access or subscriptions, Cohere optimized for three enterprise requirements that consumer-focused labs often overlooked:
1. Retrieval-Augmented Generation (RAG): Instead of relying solely on knowledge baked into model weights during training, Cohere's models were designed to retrieve relevant information from a company's proprietary databases and documents during inference. This meant businesses could get accurate, up-to-date answers grounded in their own data without expensive retraining.
"RAG fundamentally changes the economics of enterprise AI," Gomez explained at Cohere's 2023 annual conference. "You don't need a $100 million training run every time your product catalog changes. You just update your database, and the model retrieves the latest information. That's the difference between AI that's practical and AI that's prohibitively expensive."
By March 2024, Cohere's Command-R model—specifically optimized for RAG workloads—had achieved industry-leading performance on retrieval tasks while using 30% less compute than comparable models from OpenAI and Anthropic. Multiple Fortune 500 companies, including Oracle and Salesforce, integrated Command-R into their enterprise software platforms.
2. Private Deployment: Unlike OpenAI's API-based model (where customer data passes through OpenAI's servers), Cohere offered on-premise and private cloud deployment options. For heavily regulated industries—financial services, healthcare, government—this was non-negotiable. Banks couldn't send customer transaction data to external APIs; hospitals couldn't risk HIPAA violations; defense contractors needed air-gapped systems.
"We realized early that enterprises wouldn't adopt AI if it meant losing control of their data," Nick Frosst, Cohere's co-founder and Chief Technology Officer, told investors. "Our deployment model lets them run models inside their own infrastructure. We never see their data."
By 2025, Cohere's private deployment options had become a decisive competitive advantage. When JPMorgan Chase evaluated large language models for its internal code generation and document analysis tools, Cohere won the contract specifically because it could be deployed within JPMorgan's existing security perimeter. OpenAI's API-only approach was immediately disqualified.
3. Multilingual Capabilities: While OpenAI and Anthropic focused primarily on English with secondary support for major languages, Cohere invested heavily in multilingual models from the beginning. The company's Aya initiative, launched in 2023, aimed to build models supporting 100+ languages—including low-resource languages often ignored by larger labs.
This strategic choice reflected Gomez's vision of AI as a global utility rather than an English-centric technology. "There are 7,000 languages in the world," he said in a 2024 TED talk. "If AI only works well in English, we're building technology that excludes billions of people. That's both morally wrong and economically stupid."
By 2025, Cohere's Aya 23 model supported 23 languages at production quality, with Command A (released March 2025) achieving state-of-the-art multilingual performance including Arabic dialects, which most models struggled with. This made Cohere the obvious choice for multinational corporations operating across diverse linguistic markets.
Despite its technical differentiation, Cohere faced a fundamental business challenge: how to scale revenue in a market increasingly dominated by OpenAI's ChatGPT and Microsoft's aggressive bundling of AI into Office 365 and Azure.
According to financial data obtained from sources close to the company, Cohere's annual recurring revenue (ARR) reached only $13 million by the end of 2023—a fraction of OpenAI's reported $2 billion ARR and Anthropic's estimated $500 million. While Cohere had secured prestigious enterprise customers (Oracle, Salesforce, McKinsey, Accenture), deal cycles were long, contracts were often pilot projects rather than production deployments, and expansion revenue was slow.
"2023 was honestly terrifying," one early Cohere employee told us, speaking on condition of anonymity. "We had this amazing technology, great investors, strong team. But ChatGPT had completely changed customer expectations. Everyone wanted to compare us to GPT-4. Sales cycles that should have taken three months were taking nine. We were burning through cash."
Gomez responded with a three-part strategy:
First, accelerate product velocity. In 2024, Cohere shipped Command-R (March), Command-R+ (April), and Rerank 3 (September)—three major model releases in six months. Each release demonstrated measurable improvements in enterprise-critical tasks: retrieval accuracy, multilingual support, and inference speed. The rapid iteration signaled to customers that Cohere could keep pace with OpenAI and Anthropic despite having one-tenth their R&D budgets.
Second, double down on platform partnerships. Rather than competing directly with Microsoft and Google in cloud infrastructure, Cohere embedded its models into their platforms as an alternative to the incumbents' own offerings. Oracle Cloud Infrastructure (OCI) integrated Cohere models as the default LLM option in its enterprise AI suite. Salesforce offered Cohere alongside OpenAI in its Einstein AI platform. This distribution strategy turned potential competitors into channels.
Third, target specific verticals. Instead of trying to be everything to everyone, Cohere focused its sales and engineering resources on three industries where its differentiation mattered most: financial services (where private deployment was mandatory), healthcare (where multilingual patient communication was critical), and technology (where RAG for code generation and documentation was valuable). By Q2 2024, these three verticals accounted for 68% of Cohere's ARR.
The strategy worked. By January 2025, Cohere's ARR had grown to approximately $70 million—more than 5× growth in 14 months. Two sources familiar with the company's financials told us that Cohere expects to exceed $200 million in ARR by year-end 2025, with gross margins above 75% (higher than OpenAI's reported 60-65% margins, reflecting Cohere's focus on larger enterprise contracts with less price sensitivity).
On August 14, 2025, Cohere announced it had raised $500 million in new funding at a $6.8 billion valuation. The round was led by Radical Ventures and Inovia Capital (both existing investors), with heavy participation from AMD Ventures, NVIDIA, PSP Investments (one of Canada's largest pension funds), and Salesforce Ventures.
The valuation—up from $5.5 billion in its previous round 14 months earlier—came as a surprise to industry observers. In a market where AI startups' valuations were declining (Inflection AI had essentially been acqui-hired by Microsoft at a discount; Adept was struggling to raise; Character.AI had sold to Google), Cohere was commanding a significant premium.
The explanation became clear with two simultaneous announcements:
1. Joelle Pineau as Chief AI Officer: Cohere had recruited Joelle Pineau—former Vice President of AI Research at Meta and a McGill University professor widely considered one of AI's most accomplished researchers—as its Chief AI Officer. Pineau would oversee research, product, and policy, reporting directly to Gomez.
Pineau's move from Meta to a startup signaled external validation of Cohere's technical trajectory. "I spent five years at Meta working on foundation models at massive scale," Pineau told Bloomberg in an interview. "What excited me about Cohere is that they're solving harder problems—how to make models work for businesses with diverse needs, limited budgets, and strict privacy requirements. That's where the real AI impact will be."
Her appointment also addressed a persistent criticism of Cohere: that its research team lacked depth compared to OpenAI's and Anthropic's stacked rosters of former Google Brain and DeepMind researchers. With Pineau—who had overseen Meta's FAIR (Facebook AI Research) lab and published over 150 papers—Cohere now had research credibility to match its commercial execution.
2. Francois Chadwick as CFO: Cohere simultaneously announced it had hired Francois Chadwick as its Chief Financial Officer. Chadwick previously served as CFO at Uber during its 2019 IPO—one of the most complex and high-profile public offerings of the 2010s. His hiring sent an unmistakable signal: Cohere was preparing to go public.
"You don't hire an IPO CFO 18 months before you need one," one venture capital source told us. "Francois knows how to navigate S-1 filings, roadshows, and post-IPO reporting. Cohere is clearly targeting a 2026 IPO window."
On March 13, 2025, Cohere released Command A—its most advanced model and the clearest demonstration yet of the company's technical capabilities.
Command A is a 111-billion-parameter model with a 256,000-token context window—comparable in scale to GPT-4 and Claude 3 Opus, but optimized specifically for enterprise workloads. According to Cohere's published benchmarks (independently verified by researchers at Stanford's HELM evaluation framework), Command A achieves:
Perhaps most impressively, Command A requires only two NVIDIA H100 GPUs for inference—a dramatic reduction from the 4-8 GPUs typically needed for comparable models. This efficiency advantage, enabled by Cohere's proprietary optimizations to the transformer architecture, translates directly into lower operational costs for enterprise customers.
"We spent two years optimizing every layer of the stack," Nick Frosst explained in a technical blog post announcing Command A. "Model architecture, quantization methods, inference engines, memory management. The result is a model that delivers GPT-4-class performance at one-third the cost."
Early adopters confirmed the performance claims. A Fortune 100 financial services company that had been using GPT-4 for internal document analysis switched to Command A in April 2025 and reported 40% cost savings with improved accuracy on domain-specific queries. An international law firm using Command A for multilingual contract review noted that the model's Arabic support was "significantly better than anything else we tested, including GPT-4."
By mid-2025, the foundation model market had consolidated into a clear hierarchy:
Tier 1 (Consumer-Focused): OpenAI dominated with ChatGPT's 200+ million weekly active users and Microsoft's aggressive bundling into Office 365, Azure, and GitHub. Anthropic's Claude had emerged as the quality-conscious alternative, favored by enterprises concerned about AI safety and constitutional design. Together, OpenAI and Anthropic controlled approximately 58% of the enterprise LLM market (34% and 24% respectively, according to a November 2024 Menlo Ventures survey).
Tier 2 (Enterprise Specialists): Cohere, with 3% market share, led a group of enterprise-focused providers including AI21 Labs (2%) and smaller regional players. While collectively capturing less than 10% of the market, these companies were growing faster in specific verticals where their specialization provided clear value.
Tier 3 (Self-Hosted Open Source): Meta's LLaMA and Mistral AI offered free/low-cost alternatives for companies with technical resources to deploy and maintain their own models. Approximately 15% of enterprises were using self-hosted open source models by mid-2025.
Despite the challenging competitive dynamics, Cohere's growth trajectory suggested a viable path to profitability and eventual public markets:
First, the enterprise market remained massively underserved. While ChatGPT had captured consumer mindshare and Microsoft had bundled AI into its office suite, most large enterprises had not yet deployed AI at scale for their most critical workflows. A Gartner survey from May 2025 found that only 12% of Fortune 500 companies had moved beyond pilot projects to production deployments of generative AI. The remaining 88% cited concerns about data privacy, cost unpredictability, and integration complexity—precisely the problems Cohere was designed to solve.
Second, Cohere's technical differentiation was widening rather than narrowing. RAG optimization required fundamentally different architectural choices than general-purpose chatbots. As enterprises invested in proprietary data infrastructure (vector databases, retrieval systems, fine-tuning pipelines), they became locked into platforms optimized for those workflows. OpenAI could add RAG features to GPT-4, but its model architecture prioritized breadth over retrieval efficiency. Cohere's entire stack was purpose-built for enterprise RAG.
Third, the multilingual advantage created network effects. As Cohere added language support and trained models on more diverse data, it became the obvious choice for any company operating globally. OpenAI and Anthropic could match English performance, but catching up across 20+ languages would require years of data collection and training—during which Cohere would extend its lead.
"The consumer chatbot market is a winner-take-most game," one enterprise software analyst told us. "But the enterprise AI market is multi-vendor by necessity. Companies want alternatives for different use cases, deployment models, and cost structures. Cohere doesn't need to beat OpenAI overall—it needs to be the best choice for a meaningful subset of enterprise workloads. They're well on their way."
Despite Cohere's momentum, the path to a successful IPO faces significant execution challenges:
1. Revenue Growth Deceleration: While Cohere's ARR growth from $13 million to an expected $200 million over 18 months is impressive, sustaining that growth rate becomes mathematically harder at scale. Reaching $400 million ARR (likely required for a successful IPO at current valuations) would require either massive customer acquisition or significant expansion within existing accounts. Both are difficult in enterprise software.
2. Competition Intensification: OpenAI's enterprise offerings are improving rapidly. The company's June 2025 announcement of fine-tuning and private deployment options directly targeted Cohere's differentiation. Anthropic's partnerships with Amazon AWS and Google Cloud provide distribution advantages Cohere struggles to match. The competitive moat that seemed wide in 2023 has narrowed considerably.
3. Market Dynamics: Public market investors remain skeptical of AI companies after several high-profile failures (including C3.ai's stock decline and IBM Watson's struggles to monetize). For Cohere to command a premium valuation at IPO, it will need to demonstrate not just revenue growth but also a clear path to profitability—something few AI-first companies have achieved.
4. Leadership Scaling: Gomez, while technically brilliant and visionary, has never managed a company at true enterprise scale (1,000+ employees, $500M+ revenue, public company compliance). The August 2025 executive hires (Pineau and Chadwick) address some gaps, but successfully navigating an IPO and its aftermath requires experienced operators across the entire C-suite.
One former Cohere executive, speaking anonymously, expressed concerns about the company's readiness: "Aidan is an exceptional researcher and product thinker. But running a public company is different—it's about quarterly earnings, analyst expectations, institutional investor relations. I hope they're giving him the support infrastructure he needs."
Beyond Cohere's specific trajectory, Gomez's strategy represents a broader thesis about AI's evolution: that the real value won't be in consumer chatbots or AGI moonshots, but in boring, reliable infrastructure that businesses depend on daily.
"Everyone's obsessed with AGI timelines and whether AI will be sentient," Gomez said in a September 2025 podcast interview. "But the actual AI revolution happening right now is much more prosaic—companies automating document review, customer support routing, code documentation. That's where the money is. That's where the impact is."
This infrastructure-first mindset shapes Cohere's product roadmap and go-to-market strategy. Instead of chasing benchmark improvements on academic evaluations, Cohere optimizes for metrics enterprise IT teams care about: inference latency, cost per query, accuracy on domain-specific tasks, ease of integration. The result is a product that may not generate Twitter excitement but solves real problems for paying customers.
If Gomez is correct that enterprise infrastructure—not consumer applications—represents AI's largest market opportunity, then Cohere's positioning may prove prescient. The company is building toward a future where AI models are utility infrastructure like databases and cloud computing: essential, reliable, and profitable, but not exciting enough to dominate headlines.
Cohere's Toronto headquarters provides both strategic advantages and constraints. On the positive side, Canada's AI research ecosystem—anchored by Geoffrey Hinton, Yoshua Bengio, and Richard Sutton—offers deep talent pools and academic partnerships. The University of Toronto's Vector Institute (where Gomez serves on the advisory board) produces hundreds of ML graduates annually, many of whom join Cohere.
Canada's immigration policies also enable faster talent acquisition than the US. "When we need to hire a researcher from India or China, the Canadian visa process takes weeks, not months," Cohere's head of talent told BetaKit. "That speed advantage matters when you're competing with OpenAI and Anthropic for the same candidates."
However, being headquartered outside Silicon Valley creates challenges. Enterprise customers still expect their AI vendors to have substantial US presence for support and partnership. Cohere has addressed this by opening large offices in San Francisco, New York, and London, but maintaining cohesion across distributed teams remains difficult.
The 2025 funding round's heavy Canadian participation (PSP Investments, Radical Ventures, Inovia Capital) reflected both patriotic pride in a homegrown AI champion and practical investment math: Cohere represents Canada's best chance at producing an AI company with global impact comparable to OpenAI or Anthropic.
Eight years after "Attention Is All You Need," the transformer's eight authors have taken remarkably divergent paths:
Of the eight, Gomez is the only one building a pure-play foundation model company targeting enterprise markets. His trajectory—from youngest co-author to CEO of a $6.8 billion unicorn—represents perhaps the most direct commercial exploitation of the transformer's potential.
"The transformer gave us the architecture," Gomez reflected in a 2025 interview. "But architecture alone doesn't create value. You need to build products people actually want to pay for, solve real problems, and execute consistently over years. That's what Cohere is—the transformer idea translated into enterprise infrastructure."
As Cohere prepares for its IPO (likely in Q2 or Q3 2026 based on typical 18-24 month timelines after hiring an IPO CFO), the company faces critical execution milestones:
Revenue Acceleration: Reaching $300-400 million ARR by mid-2026 is essential for a successful public offering at current valuations. This requires not just landing new enterprise customers but also demonstrating expansion revenue from existing accounts—a key metric public market investors scrutinize.
Platform Partnerships: Deepening relationships with Oracle, Salesforce, and cloud providers to drive distribution at scale. Cohere's ability to become the "default alternative" to OpenAI in enterprise software platforms could determine its long-term competitive position.
International Expansion: While Cohere has offices in London, Paris, and Seoul, international revenue remains a small fraction of total ARR. Successful global expansion—particularly in Europe (where data sovereignty concerns favor Cohere) and Asia (where multilingual capabilities provide advantages)—could significantly expand addressable market.
Model Performance: Maintaining technical parity with OpenAI's and Anthropic's frontier models while optimizing for enterprise needs. The Command model family has kept pace so far, but any significant capability gap would erode Cohere's differentiation.
Team Scaling: Growing from approximately 400 employees today to 1,000+ required for a public company infrastructure while maintaining culture and execution velocity.
The AI market's volatility adds uncertainty. If consumer enthusiasm for AI wanes or enterprises slow adoption due to economic headwinds, Cohere's growth could stall. Conversely, if the enterprise AI market accelerates faster than expected, Cohere's head start in RAG optimization and multilingual capabilities could drive outsized returns.
At 28, Aidan Gomez has already achieved more than most entrepreneurs accomplish in entire careers: co-authoring one of computing's most influential papers, building a unicorn company, and positioning a credible challenger to OpenAI's enterprise dominance.
Yet those who know him describe Gomez as still fundamentally a researcher rather than a traditional CEO. He continues to write code, review model architectures, and engage in technical debates with Cohere's research team. His Twitter feed mixes company updates with arcane discussions of attention mechanism variants and training optimizations.
"Aidan is happiest in front of a whiteboard talking about model architectures," one investor told us. "The CEO stuff—fundraising, sales, press—he does because he has to, not because he loves it. His superpower is technical vision and product instincts, not operational management."
This creates both opportunity and risk. Gomez's technical depth enables him to make product decisions that differentiate Cohere from competitors led by pure business executives. But successfully navigating a public offering and managing a public company requires skills orthogonal to research—skills Gomez is still developing in real-time.
The executive team hired in 2025 (Pineau, Chadwick) suggests Gomez understands his limitations and is surrounding himself with operators who can handle the business complexity while he focuses on technology strategy. Whether this proves sufficient will become clear in the coming 18 months as Cohere moves toward public markets.
In an AI landscape dominated by OpenAI's consumer virality and Anthropic's safety-first positioning, Aidan Gomez has carved out a third path: enterprise-focused, infrastructure-oriented, and optimized for the boring but profitable work of making AI actually useful for businesses.
Whether this strategy can support a successful public offering at a $7-10 billion valuation remains uncertain. The enterprise AI market is still nascent, competitive dynamics are intensifying, and Cohere faces execution challenges that could derail its trajectory.
But if the bet succeeds, Gomez will have demonstrated something important: that the AI revolution's biggest winners won't necessarily be those building toward AGI or dominating consumer mindshare, but those solving real problems for customers willing to pay for reliability, privacy, and integration with existing infrastructure.
Eight years after co-authoring the paper that changed AI forever, the 28-year-old from Brighton, Ontario, is attempting to prove that transformers' greatest impact lies not in chatbots or artificial general intelligence, but in the unglamorous work of making enterprise software actually intelligent.
The next 18 months will determine whether that thesis can support a multi-billion-dollar public company—and whether Aidan Gomez's name will be remembered not just for "Attention Is All You Need," but for translating that academic breakthrough into enduring commercial value.