Three things happened in 48 hours that, taken together, redraw the architecture of enterprise AI.
Perplexity launched a product called Computer — a multi-model orchestration platform that coordinates 19 different AI models to execute entire projects autonomously. Nvidia posted record earnings and fell 5.5% — its worst single-day drop in ten months. And the Pentagon's Friday deadline for Anthropic to comply with unrestricted military use of Claude arrives tomorrow.
Each story looks different on the surface. Underneath, they're all about the same structural shift: the value in AI is moving from the model layer to the orchestration layer. And the enterprises that understand this will architect their AI infrastructure very differently from those that don't.
Perplexity Computer: Orchestration as a Product
On February 25, Perplexity launched what its CEO Aravind Srinivas calls the company's most ambitious product in its three-year history. It's not a new model. It's not a chatbot upgrade. It's a multi-model orchestration platform.
Here's how it works. You describe an outcome — a finished website, a competitive analysis, a data pipeline, a research report. Computer breaks that objective into subtasks and assigns each subtask to whichever AI model is best suited for it. Claude Opus 4.6 handles core reasoning and coding. Google's Gemini powers deep research. GPT-5.2 manages long-context recall and web search. Grok handles lightweight, speed-sensitive tasks. Nano Banana generates images. Veo 3.1 processes video. Nineteen models total, each doing what it does best.
Sub-agents run in parallel, each in its own isolated compute environment with browser access, filesystem access, and hundreds of tool integrations. If one agent hits a wall, the system spawns additional agents to troubleshoot — without waiting for human intervention. Projects maintain persistent memory across sessions. A workflow can run for hours or months.
VentureBeat described it as somewhere between OpenClaw (the viral open-source autonomous agent) and Claude Cowork (Anthropic's enterprise collaboration tool). The key difference: Computer operates entirely in the cloud, in a managed environment with clear governance boundaries — users set spending caps per project and can override model assignments, pinning sensitive subtasks to specific models.
From an engineering perspective, the architecture is significant for three reasons.
First, model specialisation is now the default assumption. Perplexity's thesis, stated explicitly, is that no single model excels equally at reasoning, coding, visual processing, research, and real-time search. Their own internal testing found that marketing teams produce better results with Gemini, engineering teams produce better results with Claude, and no company operates with only one type of team. The orchestration layer resolves this by routing each task to the right model automatically.
Second, the orchestration logic is the intellectual property. The models themselves are third-party — Claude, Gemini, GPT, Grok. What Perplexity built is the routing intelligence: which model handles which task, how subtasks are decomposed, how agents coordinate in parallel, how failures are handled, how context persists across sessions. That orchestration layer is the product. The models are interchangeable components underneath it.
Third, multi-model governance is built into the architecture. Users can see which model is handling which subtask, override assignments, and set per-project spending limits. The system runs in a sandboxed cloud environment rather than on local hardware. For enterprise deployment, this pattern — orchestration with visibility and control — is the governance model that scales.
Perplexity Computer is currently available to Max subscribers at $200/month, with Pro and Enterprise tiers planned. The product validates a pattern that enterprise AI teams have been building toward: the orchestration layer isn't a feature of the AI system — it is the AI system.
Why Nvidia Fell After Beating Every Estimate
Nvidia posted record quarterly revenue of $68.1 billion — beating expectations by nearly $2 billion. Guided $78 billion for next quarter — $6 billion above consensus. Data centre revenue grew 75% year-over-year. Net income nearly doubled to $43 billion. Vera Rubin samples shipped to customers. Blackwell Ultra delivers 50x performance improvement for agentic AI.
The stock fell 5.5%.
Bloomberg reported it as Nvidia's worst single-day decline in ten months. Analysts said the results “failed to dispel fears of an AI bubble.” The question Wall Street is asking: will all this infrastructure investment actually pay off?
Here's why this matters for enterprise architecture — and why the market reaction is actually the more informative signal.
The $700 billion in hyperscaler capex being deployed this year is building the compute layer. Nvidia's chips are the engines. But engines alone don't produce business value. Business value comes from what runs on top of that compute — the AI agents, the orchestration layers, the workflow automation, the systems that turn raw compute into operational outcomes.
The market's scepticism isn't about whether AI infrastructure will be built. It clearly will. The scepticism is about whether the companies deploying AI will capture returns proportional to the infrastructure cost. That's a question about the application layer, not the compute layer.
Nvidia's CFO said during the earnings call that they expect “sequential revenue growth throughout calendar 2026” and have visibility to over $500 billion in Blackwell and Rubin revenue commitments. Huang said “compute equals revenues” and “without tokens there's no way to grow revenues.” The infrastructure thesis is secure.
But the enterprise question remains: once you have the compute, do you have the orchestration layer that turns it into measurable business outcomes? The 5.5% drop is the market signalling that the infrastructure story alone isn't sufficient. The application and orchestration layer is where value gets created — or lost.
The Pentagon Deadline: Single-Model Risk in Production
Tomorrow at 5:01 PM, the Pentagon's deadline for Anthropic expires. Defense Secretary Hegseth gave CEO Dario Amodei until Friday to agree to unrestricted military use of Claude — or face contract termination, a supply chain risk designation, and potential invocation of the Defense Production Act.
Anthropic has said it won't budge on its two red lines: no mass surveillance of Americans, no fully autonomous weapons. The Pentagon has approved xAI's Grok for classified systems with no restrictions. OpenAI and Google have agreed to remove safeguards for unclassified military use.
From a pure engineering perspective, the Pentagon-Anthropic dispute is a case study in single-model architecture risk in production. Claude is the only AI model currently on the military's classified networks. A senior Defence official acknowledged that replacing it would be, in their words, “extremely painful” because Anthropic's model is that capable for specialised government applications.
The lesson isn't about military AI policy. The lesson is architectural. When your mission-critical AI infrastructure depends on a single model from a single provider, your operations become hostage to that provider's business decisions, safety policies, pricing changes, and relationship with every other stakeholder they serve.
The Pentagon is now scrambling to add providers — xAI, OpenAI, Google — specifically because single-provider dependence created unacceptable operational risk. Every enterprise running production AI workloads should ask the same question: if your primary AI provider changed terms tomorrow, what happens to your operations?
The Architecture Pattern That Connects All Three
Perplexity built a 19-model orchestration layer and turned it into a product. Nvidia built the compute infrastructure and the market asked “but what runs on top of it?” The Pentagon learned that single-model dependence creates operational fragility.
The common thread: the orchestration layer — the intelligence that selects the right model for the right task, manages governance and switching, and turns raw AI capability into business outcomes — is where enterprise value is being created.
This has direct implications for how we build AI systems for our clients.
Model Selection as a Dynamic Function
In our Minnato platform, model selection isn't a one-time vendor decision. It's a continuous function. An AI agent processing invoices might use one model for OCR extraction, another for field validation, and a third for exception handling — routing each subtask to whichever model delivers the best accuracy-to-cost ratio at that moment.
This is the same architectural pattern Perplexity Computer uses: decompose the objective, route subtasks to specialised models, coordinate the results. The difference is that Minnato applies it to enterprise operations — procurement, compliance, customer service, reporting — rather than consumer research and coding tasks.
Governance at the Agent Level
Perplexity Computer lets users set spending caps per project, override model assignments, and observe which model handles which subtask. This governance model — per-agent visibility and control — is essential at enterprise scale.
When AI agents handle sensitive operations like contract analysis, customer conversations, or financial reporting, governance can't be a binary switch. It needs to operate at the task level: which agent accesses which data, which model processes which document type, what audit trail each decision generates, and what escalation path exists when confidence drops below threshold.
Our Vult document intelligence system implements exactly this — task-level routing where invoice extraction, contract clause identification, and multilingual processing can each use the most accurate model available, with full audit trails and human escalation built in. Our Dewply voice AI does the same for customer conversations — routing speech recognition, sentiment analysis, and response generation to specialised models based on the conversation's language, emotional context, and complexity.
Infrastructure That Captures Every Improvement
Nvidia's cost curve — Hopper to Blackwell to Vera Rubin — delivers orders-of-magnitude improvements in inference cost every 12 to 18 months. But only architectures designed for model switching actually capture those improvements.
If your AI infrastructure is built around a single model from a single provider, you capture cost improvements only when that specific provider reduces prices. If your infrastructure is built around an orchestration layer that can evaluate and switch between models, you capture every cost improvement across the entire market — from every provider, every chip generation, every pricing change.
This is the fundamental engineering argument for model-agnostic architecture. It's not about hedging risk (though it does that). It's about ensuring your AI operations automatically benefit from the fastest-moving cost and capability curve in technology.
What This Week Proved
In five days:
Samsung shipped a phone with three AI agents, proving multi-model is the consumer default. Nvidia posted record earnings and the market asked “what runs on top of the infrastructure?” Perplexity launched a 19-model orchestration platform, proving that the orchestration layer is a product category. The Pentagon reached crisis point because single-model dependence created operational fragility.
The pattern is consistent. The models are becoming commodities. The compute is becoming abundant. The orchestration layer — the intelligence that selects, routes, governs, and coordinates AI models to produce business outcomes — is where defensible value is being built.
For engineering teams building enterprise AI: design the orchestration layer first. The models will keep changing. The compute will keep getting cheaper. The orchestration logic — how you decompose tasks, route to specialised models, govern agent behaviour, and maintain audit trails — is the architecture that lasts.