The Board Room

The Largest AI Capital Event This Cycle

Google sold $80 billion in new equity, its first stock issuance since 2005, to fund $180-190B in AI infrastructure capex for 2026 with 2027 set higher still. A company generating $165B in pre-capex cash flow and holding $126B in cash has concluded that operating cash flow alone cannot fund the AI buildout at competitive speed. Berkshire Hathaway anchored at $10B. Goldman Sachs and JPMorgan led the book.

When Warren Buffett's team underwrites the AI infrastructure thesis at near all-time-high prices, the bubble debate is effectively over at the institutional level.

What This Says About Everyone Else

The interesting read is not about Google's treasury decisions. It is about every company one or two tiers below. Google Cloud grew from 6% to 22% of Google Services revenue in seven years, with margins expanding to 33%. The $80B raise is a bet that Cloud's AI-driven trajectory eventually rivals the advertising business. The cloud provider most enterprises already depend on is repositioning its entire model around the assumption that customer AI consumption grows dramatically, and that customers will pay for it.

A reasonable skeptic would say capital this size has been wasted before, and the skeptic would be correct about prior cycles. What the skeptic does not explain is the structural point on which multiple sources now converge: AI competition has bifurcated into firms that can access capital at this scale and firms that cannot. There is no third category. The $80B is not being raised to lower prices for downstream customers. It is being raised to secure capacity, land, and power that competitors will not have.

The Negotiating Window Is Now

Hyperscalers racing to fill new capacity need committed demand to justify the buildout. Buyer leverage on multi-year compute contracts is at a local maximum. In 18-24 months, when this capital sits in operational data centers, the leverage inverts. Google will sell TPU capacity to Anthropic, a direct Gemini competitor, because the infrastructure economics work regardless of which model wins. The willingness to serve competitors signals where Google thinks the moat actually sits: compute access, not model quality.

The Capital Concentration Map

Total mobilized into the platform layer exceeds $500B in a single planning cycle. SK Hynix is doubling memory capacity over five years. The top memory maker plans capacity against customer commitments, not press releases.

The $500M Wake-Up Call

A single customer ran up $500 million in Claude API costs in one month. That is not an exotic failure. It is a normal agentic loop running unchecked at the throughput current models permit. The corroborating evidence is not subtle: Uber exhausted its entire annual AI budget in months, with its COO conceding on the record that ROI is hard to demonstrate; Amazon is actively suppressing internal AI usage incentives; Snowflake's CEO has said, plainly, he is "absolutely" worried about inference costs.

Inference cost behaves more like fraud exposure than like cloud compute. It can compound inside a single agentic loop in hours, not weeks, and the people authorizing the spend are usually three layers removed from the people who notice the burn rate.

The Enterprise Response: Model Routing and Cost Governance

A three-layer optimization stack is forming across the industry. Routing at the infrastructure level picks the cheapest capable model per task. Prompt engineering at the application level cuts tokens. Governance at the organizational level restricts premium model access by role. Snowflake, Palo Alto Networks, UiPath, and Zscaler arrived at variants of the same pattern independently. That is market-forced convergence, not fashion.

The numbers support the discipline. UiPath achieved 90%+ cost savings through prompt engineering alone. Novo Nordisk found that plain Excel outperformed Claude on structured clinical trial analysis. The era of indiscriminate adoption is closing. What replaces it is surgical allocation: AI reserved for tasks where it beats the alternative on a cost-adjusted basis.

The Pricing Floor Is Collapsing

MiniMax's M3 model claims near-Claude Opus 4.7 performance at 40x lower cost per million input tokens. A reasonable skeptic would discount the claim by half, and the skeptic would be right to. The skeptic still has a problem: every product priced against last year's inference costs does not survive next year. Meta is spending $125-145B on capex and signaling willingness to resell excess capacity against AWS, Azure, and GCP. Combine that with 100x annual efficiency gains in models and inference price collapse stops being a tail risk. It becomes the expected path.

The Paradox for Buyers

Per-token pricing keeps falling. Per-workload spend keeps rising. Usage is compounding faster than unit cost is falling. The line item goes up even as the unit price goes down. Vendors know this, which is why the discount sheets arriving this fall are the tell. Multi-year minimums look different when next year's effective price is a coin flip.

The board-level question has shifted. It is no longer whether the organization is using enough AI. It is whether the organization is using AI intelligently enough. The leaders winning this transition treat AI as an expensive resource to allocate precisely, with the same discipline already applied to headcount and cloud compute.

One Vendor, Every Layer

NVIDIA used Computex to declare a full-stack war. In close succession it shipped Cosmos 3 (state-of-the-art multimodal world model), Nemotron 3 Ultra (a 550B-parameter open-weight LLM scoring 48 on the AI Intelligence Index against the next-best open model at 39), and RTX Spark (a personal AI computer with 128GB unified memory and 1 PFLOP FP4). The Cosmos Coalition with Runway reads like early Android. Open enough to attract partners, controlled enough to keep the gravity where the silicon is.

When the entity that sets the price of compute is also setting the reference design for the rack, the interconnect, and increasingly the runtime, the middle of the stack is not a neutral place to build.

The Four Stakeholders Being Squeezed

1. Cloud providers now compete with NVIDIA on model hosting against open-weight alternatives tuned for NVIDIA silicon
2. Model providers face Nemotron 3 Ultra delivering near-frontier performance as free open weights, which erodes pricing power directly
3. Hardware companies (Apple, Intel, AMD, Qualcomm) must defend consumer and edge AI against RTX Spark's specs
4. Developer tools companies face a full-stack ecosystem with enough gravitational pull to marginalize point solutions

ARM moved 16% in a single session. Intel and AMD face existential questions. Qualcomm's Snapdragon X bet gets harder to defend. The N1X chip, partnered with Microsoft for the fall 2026 Windows launch, positions NVIDIA silicon as the standard inference layer in every endpoint device.

Local 120B Models Become Practical

RTX Spark running 120B-parameter models locally with full CUDA support means that within 18 months, competitors will run capable models at zero per-inference cost, with full data privacy and sub-millisecond latency. Every product roadmap built on the assumption that cloud inference is the permanent delivery mechanism needs a second pass. The cloud-to-edge shift is no longer theoretical.

The Strategic Tradeoff

Building closer to the NVIDIA stack accelerates everything in the next 12 months and concentrates the failure mode in the next 36. Building away slows the next 12 and preserves optionality that may or may not be worth the cost. There is no third option that is honest about both timelines. The worst position, which is where most platforms actually sit, is to have picked the NVIDIA side already without admitting it on a slide.

The agent orchestration layer is still contestable, which is where the next two years of differentiation lives. The window to build differentiated agent infrastructure narrows as Google, NVIDIA, and others ship managed solutions that will be good enough for most use cases within 2-3 quarters.

The Category Redefinition

Anthropic shipped Dynamic Workflows, a 1,000-agent orchestration system that ported the Bun runtime, 750,000 lines of code, in 11 days at a 99.8% test pass rate. The board-deck version is that this is another coding-assistant release. The complete version is that it is a replacement for a mid-sized implementation team. Standard token pricing held flat at $5/$25 per million while fast mode dropped 67%. Anthropic is pricing to be picked.

Microsoft used the same week to confirm that Mustafa Suleyman's MAI team is building a coding model explicitly to "take GitHub Copilot back from Claude Code." The OpenAI partnership is functionally over as an exclusive arrangement. OpenAI extended Codex to operate any Windows application via Computer Use. What looked like a two-firm alliance last quarter is a three-firm contest this quarter.

The Infrastructure Is Buckling

GitHub reports a 1,400% year-over-year jump in coding-agent activity, 275 million commits per week, and Actions compute doubled to a billion minutes weekly. The interesting thing about the bottleneck is that it is not bandwidth. It is the permissioning layer, MySQL One, a monolithic database that was never designed to authenticate non-human reasoning at machine speed. GitHub needs CPUs, not GPUs. The unglamorous middle of the stack, between agent output and production code, is where the next twelve months of pain lives.

The code-generation problem is solved. The trust-verification problem is not. Every organization deploying AI coding tools is producing more output, more bugs, more churn, and net-negative productivity once rework is counted — unless they've rebuilt around orchestration and verification.

The Organizational Gap

The dissonant finding across sources is the one that matters. individual developers are measurably faster with AI assistance, while organizations are not shipping more business value. The gap is structural. AI optimizes the part of the lifecycle that was already cheap, which is writing first drafts. The bottleneck, reviewing and integrating and testing against systems nobody fully understands, did not move. Cheaper drafts produce more drafts, which produce more review load. Amazon claims 4,500 developer-years saved. The METR/Anthropic RCT shows a 19% slowdown. Meta sees one new insecure AI integration per week. Both sets of numbers are real. Only one makes it into the strategy memo.

The Security Dimension

Claude Mythos found 23,019 security vulnerabilities in open-source code in its first month. AI agents accessed 1,500 unauthorized database tables inside Meta. Slopsquatting, where models hallucinate package names that attackers then register with malicious code, is now recurring rather than anecdotal. A codebase where 46% of commits are AI-generated, trending toward 60%, is one where conventional review practices stop scaling. The decision this quarter is whether the review architecture is rebuilt now or rebuilt after the incident.