The End of the Single-Vendor AI Dream
Multi-model AI routing is an infrastructure strategy where enterprises send each AI request through a central gateway that dynamically chooses among several large language models based on price, performance, and task fit, instead of locking into a single provider for all workloads.
The era of picking one favorite AI lab and building everything on top of it is over. Vercel CEO Guillermo Rauch says companies are no longer relying on a single AI lab and calls one-lab partnerships obsolete. This is not a theoretical prediction; it is coming from a platform that now routes more than a trillion tokens a day across millions of deployments. When the person running that much traffic says loyalty is dead, believe it. Rauch argues that teams now understand their AI stack—model, harness, data platform, sandbox, and gateway—and every piece is plug and play. You can use OpenAI, Anthropic, Gemini, or Chinese models like GLM-5.2 and GLM-5.2’s peers as interchangeable parts.
The reason is blunt: enterprise AI spending is no longer a vanity metric. Executives are realizing that money spent on AI is not directly translating to more value shipped for customers. The days of telling employees to burn as many AI tokens as they can are over; now the mandate is AI cost optimization through smarter infrastructure, not bigger model contracts.
Vercel’s Bet: Models Are Commodities, Gateways Are Strategy
Rauch’s most important claim is not that Gemini or any specific model is winning, but that price-to-performance now beats brand. He says Gemini is seeing strong growth because its models offer “awesome price/performance characteristics” when scaling up. The message is that a frontier model is no longer a precious snowflake; it is a swappable component inside an inference pipeline.
In this view, the AI model gateway strategy becomes the real control plane. Instead of binding applications directly to a provider, companies front them with a gateway that can route work across multiple models from different labs. Like a centralized control plane, that gateway intercepts every prompt and makes a split-second decision: does this request need an expensive frontier model or can a cheaper alternative handle it just as well? When you are routing over a trillion tokens per day, those decisions compound into millions saved—or wasted—on enterprise AI spending.
This is the quiet but radical shift: the model is no longer the product. The surrounding infrastructure—routing logic, evaluation pipelines, caching layers—is where differentiated strategy now lives. Trying to pick the single best AI provider is a losing game when capabilities converge and prices diverge.
Inside Coinbase’s 50% AI Cost Cut
If Vercel describes the new architecture, Coinbase proves it with a line item. Coinbase CEO Brian Armstrong reports that the company cut its internal AI spend by nearly half while overall token usage continued to grow, without imposing usage caps on engineers. That result did not come from begging staff to type less; it came from multi-model AI routing and disciplined defaults.
Coinbase runs its own internal LLM gateway and deliberately defaults engineers to lower-cost open-weight models like Z.ai’s GLM 5.2 and Moonshot AI’s Kimi 2.7. GLM 5.2 is priced at roughly USD 1.40 (approx. RM6.40) per million input tokens and USD 4.40 (approx. RM20.00) per million output tokens, compared with Anthropic’s Opus 4.8 at around USD 5 (approx. RM23.00) for input and USD 25 (approx. RM115.00) for output. That three- to six-times cost reduction per token, paired with competitive coding benchmark scores, makes the cheaper models the sensible default.
On top of cheaper defaults, Coinbase uses task-based model routing: frontier models do complex planning, while simpler execution work goes to cheaper models where performance is similar. Aggressive caching then keeps conversations pinned to the same model so context can be reused; by doing this, Coinbase pushed its cache hit rate from about 5% to 60%, a twelve-fold improvement that became a major cost driver.
1,200 AI Agents and a Human-Free Routing Layer
Armstrong’s most provocative idea is that humans should not be picking models at all. Coinbase now operates with roughly 1,200 full-time AI agents, measured by normalizing compute hours to a 40–60 hour workweek. At that scale, he argues, expecting human developers to choose the right model for each task is a waste of talent and an invitation to inconsistent costs.
Instead, the gateway acts as the automated brain: it inspects the prompt, checks cache state, looks at real-time pricing, and decides which provider and model to use. Multi-model AI routing becomes as central to engineering as API gateways were to web services. But that power comes with a new demand for observability. Teams now need visibility into latency, uptime, token consumption, and cost across all providers to know whether routing choices are improving performance or quietly bleeding money.
Evaluation pipelines are also non-negotiable. Lower-cost models must be continuously tested against an organization’s own workloads before they get production traffic, because public benchmarks cannot tell you how they will behave on your specific code, data, and workflows. In other words, multi-model infrastructure only works if you invest in measurement as seriously as you invest in models.
From Monolithic AI Deals to Modular Infrastructure
The Vercel and Coinbase stories point to a broader enterprise turn away from monolithic AI subscriptions and toward flexible, modular AI infrastructure. Last year, many companies were in prototyping mode, building agents without worrying too much about cost. Now they are living with agents in production, confronting the gap between AI spending and customer value, and restructuring their stacks in response.
Rauch’s observation that every component of the AI stack is plug and play—from model to gateway—captures the new reality. Both Vercel and Coinbase assume that today’s top model will not stay on top for long, so competitive advantage shifts away from the models themselves and toward the routing and control layer that decides which models to use. Multi-vendor strategies let companies route each request to the most cost-efficient model for its task, squeezing far more value out of every token.
The lesson is clear: enterprises that cling to single-vendor AI deals will pay a premium for no good reason. Those that treat models as interchangeable parts, invest in AI cost optimization, and build smart gateways will not only slash their bills; they will move faster the next time the model leaderboard changes. In an environment where frontier models keep converging on capability while diverging on price, multi-model AI routing is not a nice-to-have—it is the new baseline for serious AI adoption.
