AI & Tech Brief ⚡

📌 Navigate

📊 Exec Summary

The week of Feb 9-15 crystallized two concurrent stories: the US-China AI conflict escalated from a technical dispute to a congressional geopolitical event, while the enterprise agentic platform race opened with OpenAI's Frontier launch — both forcing every frontier lab to pick a side on architecture, governance, and infrastructure access.

Five things moved in AI/tech this week:

The pattern: geopolitics and product are no longer separate tracks — this week they collided.

1️⃣ OpenAI's DeepSeek Memo to Congress

TL;DR: On Feb 12, OpenAI sent the House Select Committee on China a memo alleging DeepSeek employees are using obfuscated third-party routers to circumvent access controls and distill frontier US AI models — framing the practice as a national security threat.

What happened

• OpenAI submitted a formal memo to the House Select Committee on China on February 12, 2026
• Allegation: accounts linked to DeepSeek employees using obfuscated third-party routers to conceal origin and access OpenAI models
• Technique described: "distillation" — using outputs of a larger model to train a smaller proprietary one that acquires similar capabilities
• OpenAI stated that distillation activity linked to China (and occasionally Russia) is becoming "more covert and complex" despite ongoing crackdowns
• National security framing: concern that "authoritarian governments deploy frontier AI for offensive cyber operations, disinformation campaigns, and mass surveillance"

🔗 Primary source → OpenAI Accuses DeepSeek of Distilling US AI Models — Bloomberg

🔍 The non-obvious point

The congressional framing is the load-bearing choice. By routing this through the House Select Committee on China — not courts, not regulators — OpenAI made this a foreign policy instrument, not an IP dispute.

• This creates legislative pressure on AI export controls beyond chips, potentially encompassing model access, API restrictions, and API-layer security requirements
• Anthropic's later Feb 23 allegations are later context, not week-native evidence; they still point to the broader lobbying strategy across labs
• For AI platform builders: model access governance is now a compliance surface, not just a business concern

👀 What to watch

• Congressional response to the memo — any markup or hearing scheduled for Q1 2026 will determine whether distillation becomes a regulated practice

2️⃣ Nvidia H200 Licensing Review Revised for China (supply-chain / export control — confirmed policy action, not shipped product)

TL;DR: On Jan. 13, BIS revised license-review policy for Nvidia H200 AI chip exports to China on a case-by-case basis if security requirements are met — a policy that sits in the same broader export-control window as OpenAI's later congressional memo.

What happened

• BIS revised H200 export license-review policy on January 13, 2026
• The policy covers H200, AMD MI325X, and similar chips on a case-by-case basis if security requirements are met
• Sales implications remain tied to case-by-case review rather than a blanket tightening date in this week's reporting
• Rep. Gallagher wrote to the Commerce Secretary: "Every advanced AI chip we send to China accelerates the PLA's development of autonomous weapons systems and surveillance capabilities"
• Trump administration had in December allowed H200 sales provided the US received a 25% cut of revenue

🔗 Primary source → Department of Commerce revises license review policy for semiconductors exported to China

🔍 The non-obvious point

The BIS revision and the Feb 12 congressional memo are the same political event playing out across two policy tracks simultaneously.

• Chips and model access are converging into a single export control regime — one restricting hardware, the other restricting intelligence
• For companies building AI infrastructure dependent on Nvidia supply chains: licensing volatility is now a planning variable
• The distillation + chip control combo is designed to close both vectors of Chinese AI capability acquisition

👀 What to watch

• Commerce Department decision on H200 licensing scope — any broadening beyond case-by-case review to AMD or other chips would signal a larger supply chain freeze

3️⃣ OpenAI Frontier: Enterprise AI Agent Platform

TL;DR: OpenAI launched Frontier on February 5 — an open-architecture enterprise platform for building, deploying, and managing AI agents — with Uber, State Farm, Intuit, and Thermo Fisher among initial customers; broader availability expected over the following months.

What happened

• OpenAI Frontier announced February 5, 2026
• Open architecture: compatible with agents from OpenAI, Google, Microsoft, and Anthropic — explicit multi-vendor positioning
• Core capability: connects enterprise data silos (CRM, ticketing, data warehouses, internal apps) and gives agents shared business context
• Enables long-horizon agentic tasks including file handling, code execution, tool use, and multi-step reasoning
• Initial customers: Uber, State Farm, Intuit, Thermo Fisher Scientific
• Pricing not disclosed; broader rollout planned over coming months

🔗 Primary source → Introducing OpenAI Frontier — OpenAI

🔍 The non-obvious point

The open architecture decision is the strategic bet. By supporting agents from Anthropic, Google, and Microsoft, OpenAI is positioning Frontier as the enterprise operating layer — not a walled garden.

• This directly threatens enterprise SaaS incumbents (Salesforce, Workday, ServiceNow) whose moats depend on data-layer lock-in
• Thermo Fisher Scientific as a launch customer is notable: it's a life sciences and lab workflow company, signaling AI agent penetration into regulated, complex operational environments
• The governance and permissions layer — not the agents themselves — is where Frontier's durable value lives

👀 What to watch

• Pricing disclosure expected with broader rollout — will determine whether Frontier is positioned as a premium add-on or infrastructure replacement

4️⃣ GPT-5.3-Codex vs. Claude Opus 4.6 — Simultaneous Launch

TL;DR: OpenAI and Anthropic launched their respective agentic coding flagships within minutes of each other on February 5, signaling that agentic software development is the current primary enterprise battleground.

What happened

• GPT-5.3-Codex (OpenAI) and Claude Opus 4.6 (Anthropic) both launched February 5, 2026
• GPT-5.3-Codex: 25% faster than predecessor; full software lifecycle support (debugging, deploying, monitoring, PRD writing, user research, metrics); long-running tool-use task execution
• Claude Opus 4.6: intelligence layer for Claude's multi-agent "team" features; available on claude.ai, API, and major cloud platforms

📊 Benchmarks (from public announcements)

🔗 Primary source → OpenAI launches new agentic coding model only minutes after Anthropic drops its own — TechCrunch

🔍 The non-obvious point

The simultaneous launch is a coordination signal, not a coincidence. Both labs are racing to anchor the agentic coding category before enterprise buyers lock in platform decisions.

• OpenAI's framing (full software lifecycle) positions GPT-5.3-Codex as a developer platform replacement
• Anthropic's framing (team orchestration) positions Opus 4.6 as the coordinating intelligence layer for multi-agent workflows — less direct, more architectural
• Builders choosing now are making infrastructure bets, not just tool bets

👀 What to watch

• First enterprise adoption data from Frontier early customers — will reveal which model wins on agentic coding tasks in production environments

5️⃣ Hyperscaler Capex: $690B and Inference Primacy

TL;DR: The five largest cloud and AI infrastructure providers committed $660–690B in 2026 capex, while inference has overtaken training as the dominant AI workload and cost center for at-scale teams.

What happened

• Hyperscaler 2026 capex commitments: Amazon $200B, Google $175–185B, Meta up to $135B, Microsoft (not separately broken out), Oracle $50B (136% increase over 2025)
• Inference has overtaken training as the dominant AI workload for most teams operating at scale
• Inference is now the single biggest cost driver for at-scale AI deployment
• Structural constraints reshaping capacity: fiber networks, renewable energy procurement, hyperscaler capital allocation patterns
• Ramp data: customers spent $260M in Q4 2025 on AI infrastructure businesses — 60% of what they spent directly querying closed-source foundation model providers

🔗 Primary source → AI Capex 2026: The $690B Infrastructure Sprint — Futurum Group

🔍 The non-obvious point

The inference primacy shift means the cost structure of AI products has permanently changed — training is amortized, inference is operational.

• For AI product builders: inference cost optimization (batching, caching, model routing) is now a P&L variable, not a technical concern
• The fiber/energy constraint means geographic distribution of capacity is determined by infrastructure availability, not just chip procurement
• Benchmark gaming (some benchmarks have 42% error rates; models trained on benchmark data) is making inference performance evaluation harder precisely when inference cost matters most

👀 What to watch

• Meta's Nvidia Grace CPU deployment — first major standalone CPU use in AI data centers signals diversification away from pure GPU dependence

📊 The pattern

This week, the US-China AI conflict and the enterprise agentic platform race ran on parallel tracks and intersected at the same policy moment. Export controls tightened on chips (Feb 11), OpenAI escalated to Congress on model theft (Feb 12), and two frontier labs launched agentic coding platforms within minutes of each other (Feb 5) — all while $690B in infrastructure capex locked in inference as the primary cost center. The geopolitical and product layers are no longer separable: the same week that reshaped enterprise AI architecture also reset who gets access to the infrastructure underneath it.

👀 Watchlist

📎 Sources

Sources of truth

Also consider reading