The Board Room

The Capability Jump

Tuesday's briefing put AI offensive capability at 81% success on individual tasks. This week's data is a category change, not another tick on the curve. AISI confirmed Anthropic's Mythos as the first model to clear both simulated attack ranges end-to-end. Full network takeover, not persistence and not lateral movement. OpenAI's GPT-5.5-cyber completed one of the two. The trend line in AI cyber task completion was already doubling every few months. These results broke above that trend.

The security model of the defensive stack was built on the premise that the cost of understanding the agent exceeded the value of bypassing it for most adversaries. That premise is no longer true for a growing share of the threat population.

EDR Is Now a Glass Box

TrustedSec ran LLMs against five commercial EDR products and found all five share the same architectural pattern: YARA-style rules, behavioral logic, allowlists, prefilters, Lua scripting engines readable after a single decryption pass, and local ML classifiers. Work that took a skilled reverse engineer weeks now takes days with AI assistance. The entire endpoint security category has been running on security-through-obscurity. The obscurity just left the building.

The Exploit Window Has Collapsed

PraisonAI was actively exploited 4 hours after disclosure. Microsoft's MDASH system found 16 exploitable flaws in a single Patch Tuesday cycle using multi-model AI analysis. An 18-year undetected RCE in NGINX's rewrite module is the other half of the story: foundational infrastructure harbors defects nobody audited. AI-powered vulnerability discovery, with Mozilla finding 271 real bugs in Firefox using custom harnesses, is compounding the problem from both sides at once.

The Market Response

OpenAI's Daybreak launched with CrowdStrike, Palo Alto Networks, Cisco, Cloudflare, Oracle, Zscaler, Akamai, and Fortinet as partners. Congress is holding closed-door Mythos demos, with access routed through NSA over CISA. That ordering prioritizes offensive and intelligence operations over civilian defense, which means the private sector is on its own for several years. Security equities are up 20% YTD. The market is starting to price this in and is almost certainly underpricing it.

The Foxconn Compound

Nitrogen ransomware exfiltrated 8TB of confidential designs from Apple, Intel, Google, and Nvidia through a single contract manufacturer. Separately, a Raspberry Pi honeypot dressed as an AI stack was indexed by Shodan in 3 hours and absorbed 113,000+ attacks in a month. A reasonable skeptic would say these are two different stories. They are not. Supply chain custody risk and AI infrastructure exposure are the same asset class under two custody regimes, and both are being probed at machine speed.

The Collision

Four of the largest platforms in enterprise software moved in the same week to own the surface where AI agents commit writes, which is the execution layer between intelligence and action. SAP, ServiceNow, Apple, and Google did not coordinate. They did not need to. When four incumbents announce the same thing in the same quarter, the announcement is the market telling you the UI-centric era is closing. The consequence for any technology leader is that the product story written eighteen months ago does not survive contact with this quarter.

Agents that act across finance, HR, IT, and procurement need one authoritative place to reconcile state. Two authoritative places is zero authoritative places.

The Architectural Split

ServiceNow adopted MCP (Model Context Protocol) servers as the standard for its headless Action Fabric, which is what an incumbent does when it wants the rest of the ecosystem to settle the protocol fight on its preferred terms. SAP is playing the opposite hand, building a vertically integrated Knowledge Graph behind a €100M fund whose entire purpose is to make SAP's own agents contextually superior inside SAP's own data universe. A reasonable skeptic would say one of these companies has read the room and the other has not. The reasonable skeptic is being too tidy. These are competing theories of where value sits in the agent economy, open interoperability vs. data-moat integration, and both have worked before in adjacent fights.

Platform Gatekeepers Move

Apple is inserting itself at the agent layer on iOS, specifically at the case where agents 'spin up smaller apps on the spot after Apple has already approved the parent app.' That sentence is the entire App Store rent-extraction model translated into the agent era, which is to say it is not new behavior, just a new surface. Google's Gemini Intelligence ships this summer across more than three billion devices, which converts 97%+ market share in key markets into a default that competitors will spend years trying to dislodge. The app stops being the product, and the agent becomes the interface.

Where Value Migrates

The a16z TAM Claim

Andreessen Horowitz staked a public position this week that $150B+ of GTM value migrates from the traditional CRM toward the AI orchestration layer. The Lemkin numbers are the actual argument: 80% fewer human seats and 83% higher total spend with 20+ agents in the loop. Seat-based pricing is structurally breaking at the same time per-customer revenue is going up, which is the pattern that usually takes pricing committees four quarters to admit is real. The CRM stops being where the work happens. It becomes where the work is recorded.

Vercel's production telemetry already shows agentic workloads at 59% of all token volume, which is not a forecast, it is a Tuesday. Products without agent-compatible APIs get bypassed the way sites outside the default search index got bypassed in 2005, and most of those sites spent the next three years convinced they were a special case.

The xAI Signal

Elon Musk, who publicly called Anthropic "misanthropic and evil," has agreed to lease them 220,000 GPUs — 45% of Colossus 1. The financial logic has beaten the competitive logic. Grok never achieved meaningful traction and trails open-source models in developer surveys, and the lease revenue almost certainly exceeds what Grok could generate from the same silicon. A reasonable skeptic would call this a one-off. The skeptic should explain why one of five frontier labs is willing to financialize its infrastructure at all. The population of viable frontier labs is contracting, and the excess is moving onto the lease market.

GPU supply has become a financial instrument first and a strategic moat second. That is good news for infrastructure buyers who want optionality and bad news for anyone who told a board last year that vertical integration of compute was the durable advantage.

Power as Platform Business

Fervo Energy's IPO at a $10B+ valuation, with shares jumping 33% above an already-raised price target, prices AI power supply as a platform business rather than a utility commodity. The more useful number is Google's option for 3 gigawatts against only 658MW currently contracted. At roughly 50MW per large data center, 3GW is sixty-plus facilities from a single supplier. Power contracts signed this year set competitive position in 2028 to 2030.

The Lockup Accelerates

Court filings put Microsoft's commitment at $100B+ to OpenAI by June 2026, against $30B in direct revenue. OpenAI separately committed $20B to Cerebras, underwriting its $56B IPO, which popped 70% on day one. Nebius reports 4+ customers per GPU while growing 684% year-over-year. The supply curve for frontier silicon is being negotiated in private, in blocks of $10B and up.

What This Means for Enterprise Buyers

The optionality most infrastructure plans were quietly relying on, the assumption that capacity would be available somewhere at some price when the workload showed up, is the line item being deleted. The xAI lease and the broader financialization of GPUs could meaningfully alter compute economics for enterprises over the next 12 to 18 months as excess infrastructure hits the secondary market. The two trends sit in tension: long-term capacity is being locked up by hyperscalers, while short-term supply may improve as failed frontier players become landlords. The decision this quarter is which side of that tension a buyer underwrites.

Infrastructure Investment Hierarchy

1. Frontier compute, locked via bilateral commitments such as OpenAI/Cerebras and Microsoft/OpenAI
2. Power supply, being platformized via Fervo/Google at 3GW
3. Neocloud capacity, with 4:1 demand and 684% growth at Nebius
4. Secondary market, emerging via the xAI/Colossus lease

The Blueprint

a16z has published what is, on any honest reading, the most comprehensive AI liability lobbying blueprint the industry has produced. The headline proposals are user-liability defaults and damages caps. The subtext is that the venture class has decided the legal architecture of the next decade is worth $115.5M in political capital right now, which makes a16z the largest disclosed donor of the 2026 midterm cycle. Marc Andreessen sits on the White House tech council.

The Tempo Mismatch

Courts are deciding cases now. Congress is debating. The likely sequence is precedent-setting rulings landing before any comprehensive federal framework, producing a patchwork of judicial standards that subsequent legislation has to work around rather than replace. Active litigation against general-purpose AI tools could impose substantial penalties on developers for user misuse before a legislative framework exists at all.

The competitive moat for any serious operator in this space for the next five years will be the quality of the audit trail, the defensibility of the evaluation process, and the contractual allocation of residual risk with upstream vendors.

The Open-Source Threat

If developer liability for downstream use becomes the standard, the economic logic of releasing an open-source model stops working. No rational actor open-sources a model that generates unbounded liability for every downstream application. Product strategies that quietly assume continued access to open weights, which describes most of them, carry an unpriced dependency on regulatory outcomes that the P&L does not show.

The ODNI vs. Commerce Battle

Inside the White House, the intelligence community is proposing a center inside ODNI for evaluating new AI models before release. It is a licensing regime in everything but name. Commerce's alternative, voluntary agreements through CAISI, preserves speed-to-market but lacks enforcement. CAISI published and then retracted voluntary testing agreements with Google, Microsoft, and xAI in the same week. The outcome determines whether frontier AI companies operate under pre-release gating or a voluntary framework.

Two Scenarios, Different Businesses