The Board Room

Deep dives

1. 01

   OpenAI's Consulting Arm Just Compromised Your Advisory Relationships — Act This Quarter

   The Structure That Changes Everything

   OpenAI launched The OpenAI Deployment Company, or DeployCo, at a $10B pre-money valuation with four billion dollars in PE funding. The investor roster is the strategic story. TPG leads, with Advent, Bain Capital, and Goldman Sachs writing checks, while McKinsey, Bain & Company, and Capgemini came in as implementation partners. Those three consulting firms now earn a guaranteed 17.5% return for steering their Fortune 500 client base toward OpenAI's deployment apparatus.

   The acquired forward-deployed engineering team from Tomoro supplies execution capacity on day one. This is not a joint venture that quietly dissolves in eighteen months. It is a permanent structural realignment of the enterprise AI advisory market.

   ---

   Why This Is Different From Prior Platform-Services Plays

   A reasonable skeptic would point out that cloud vendors have sold professional services for years. The skeptic is correct about the precedent and wrong about the mechanism. In prior cycles the consultant and the platform were separate entities with separate incentives, and the consultant could plausibly recommend AWS, Azure, or GCP on client fit. That neutrality is now structurally compromised for three of the most trusted names in enterprise strategy.

   When the model vendor is also the implementation partner, and the strategy consultant holds equity in both, the phrase "vendor-neutral advisory" needs a new definition.

   The second-order effect matters more. OpenAI has effectively concluded that model access alone is commoditizing, and that the margin has moved downstream into transformation work. That concession, dressed as expansion, tells you API access will keep repricing toward zero while the integration layer keeps the premium. Every "we help you implement AI" startup and every mid-tier consultancy is now staring at the most formidable competitor it will meet this decade.

   Cross-Source Tension Worth Noting

   Sources diverge on timing. Enterprise AI deployment is described in the same week as a $4B addressable market ready to be captured and a market where zero of fifty Midwest CIOs have agents running at scale. Both are true. OpenAI is building the deployment engine for a market that has not arrived yet, which is the classic platform move of being ready before the customer is. Firms that build internal capability now keep optionality. Firms that wait will rent it from a vendor whose interests diverge from theirs.

   The Conflict Disclosure Problem

   Any engagement with Bain, McKinsey, or Capgemini on AI strategy now carries an embedded financial incentive. That does not make the advice wrong. It means the conflict belongs on the first page of every engagement letter and inside every board presentation where these firms sit at the table. The firms that surface it early keep their trust. The firms that get caught surfacing it in a procurement review lose credibility with the boards they spent a decade earning.

   What to do

   1. →

      Audit all active consulting engagements with McKinsey, Bain & Co., and Capgemini for AI-related scope — require written conflict disclosure by end of month

      NowThese firms are now equity investors in your potential AI vendor. Every recommendation they make carries financial incentive that must be disclosed.
   2. →

      Make the build-vs-buy decision on internal AI deployment capability within 90 days — scope a 5-person forward-deployed AI engineering team as the minimum viable alternative to DeployCo dependency

      This sprintThe window before DeployCo achieves critical mass and becomes the path of least resistance is approximately six months. After that, organizational inertia favors outsourcing.
   3. →

      Evaluate Anthropic, Google, and Mistral deployment partnerships as deliberate counterweights — request proposal from at least one non-OpenAI implementation option before any new AI project kicks off

      This quarterVendor diversification is a time-decaying asset. Every quarter of single-provider dependency raises switching costs.
2. 02

   The Compute Market Just Split in Two — Your Procurement Decision This Quarter Defines 2028 Margins

   The Bifurcation Thesis

   The AI compute market is splitting into two workloads that happen to share a name and almost nothing else, and most procurement plans are still buying for one of them. Answer inference is human-facing and latency-critical, so it stays GPU-bound for the obvious reason that users notice a half-second of hesitation. Agentic inference runs in the background, across multi-step loops, with no human waiting on the other end. It is memory-bandwidth-bound, and it does not need cutting-edge silicon to do its job.

   An agent loop running in the background against a memory-bound task should not be priced at frontier-GPU rates. The procurement plan should stop pretending both workloads need the same chip.

   Agentic inference is on track to become the largest compute market by far. It rewards cheap memory, older-node chips, standard DRAM, and CPUs, which is exactly the profile China can still build under export controls. It puts structural pressure on Nvidia's premium positioning across the majority of future demand. And it means capacity committed this quarter locks in a cost structure tuned for the smaller of the two workloads.

   ---

   Nvidia's $40B Response

   Nvidia deployed over $40 billion in equity investments in 2026 alone, and the purpose was not diversification. The purpose was making sure every layer of the AI stack runs on its hardware. This is supply-chain financing at a scale that creates lock-in at a layer above chip supply. Once Nvidia has financed the cloud provider, the training platform, and the inference vendor, moving to AMD or custom silicon stops being a technical call and becomes a relationship one.

   Nvidia itself is already behaving like a firm that sees the threat. Dynamo disaggregates inference. Standalone memory and CPU racks are shipping. That is a company preparing to compete on volume, not only on premium. The enterprise buyer's current procurement posture is financing that transition on Nvidia's behalf rather than building the buyer's own optionality.

   The Evidence From This Week

   Signal	Implication
   Anthropic secured 220,000 GPUs from SpaceX's Colossus 1	Agentic workloads consuming unprecedented compute; secondary GPU market forming
   OpenAI + Meta committed 12 GW to AMD Instinct	Nvidia monopoly broken for inference; HBM supply absorbing fast
   Mistral 20x ARR growth to ~$1B on sovereign multi-vendor pitch	Enterprises paying premium to avoid single-vendor concentration
   Google Decoupled DiLoCo: training across regions on 2-5 Gbps internet	Geographic concentration of training compute no longer required

   The Contrarian Read

   A reasonable skeptic would point out that GPUs have absorbed every workload thrown at them for a decade, and that confident predictions about "the largest compute market" have a poor track record. The skeptic is correct about the history. The history does not explain why background agent loops should pay frontier-GPU rates indefinitely. The thin-abstraction approach — routing workloads across two or three backends by shape — is the ten-percent-harder architecture that most teams will skip. This quarter's procurement sets next year's margin structure.

   What to do

   1. →

      Commission infrastructure architecture review separating 'answer inference' and 'agentic inference' workload planning with distinct hardware strategies for each — deliver findings within 60 days

      This sprintLocking in premium GPU capacity for memory-bound agentic workloads is paying for capability you won't use. The cost differential compounds quarterly.
   2. →

      Negotiate flexibility clauses into any Nvidia procurement commitments made this quarter — specifically exit ramps and workload-based pricing tiers

      This sprintNvidia's $40B financing strategy is designed to lock customers in. Flexibility now costs a small premium; flexibility later costs multiples more.
   3. →

      Evaluate AMD Instinct MI450 and commodity memory architectures for agentic workload pilots — initiate conversations before the HBM supply window narrows further

      This quarter12 GW committed by just two companies is absorbing available supply. Mid-tier buyers without locked-in capacity will face cost assumptions that are already wrong.
3. 03

   The Org Chart Is Now the Strategy: What Separates Ramp's 25% Premium From Parker's Bankruptcy

   Five Companies Converged on the Same Design This Quarter

   Ramp, Block, Coinbase, Chime, and Cloudflare independently arrived at the same organizational architecture this quarter: restructure around AI agents as primary workflow executors, flatten hierarchies, eliminate pure management roles, use the resulting efficiency to justify premium positioning. Five firms reaching the same design in the same ninety days is not a trend piece. It is convergent evolution under identical selection pressure.

   The market is pricing it already. Ramp moved from $32B to $40B+ in six months on a $1B revenue base. That is a 25% valuation step tied directly to AI-agent-embedded positioning. Cloudflare cut 1,100 employees, close to 20% of the company, named AI tooling as the reason out loud, and the stock is up 30% year to date. Block shipped Moneybot and Managerbot. Coinbase cut 700 heads, capped layers at five below the C-suite, and replaced pure managers with player-coaches.

   ---

   The Paradox: 80% of AI Layoffs Produce Zero ROI

   Set against the winners, one finding deserves a line on the board deck: eighty percent of companies that cut jobs for AI saw no improvement in returns. The difference is sequencing. The winners deployed agents, measured the gain, then restructured around proven capability. The losers treated headcount reduction as the strategy rather than the consequence of one.

   AI amplifies existing organizational capability. It does not create new capability. Well-documented, fast-moving organizations with clear workflows will compound. The rest will get marginal gains and substantial frustration.

   What the Winners Are Actually Doing

   • Sendbird: the CEO runs internal token leaderboards and does 1:1s with non-adopters. Adoption is treated as a product with engagement mechanics, not a memo.
   • Notion: rebuilt around spec-driven development, where 4-sentence Markdown prompts generate PRs. The CI pipeline is being cut 75% specifically to unlock agent throughput — 20 iterations per hour against the old 1.
   • Cloudflare: named AI tooling as the explicit reason for the 20% cut. That gives every peer CEO public permission to have the same conversation next quarter.

   The Cautionary Case

   Parker filed Chapter 7 despite raising $200M. The failure mode is familiar. Scale on favorable market conditions, skip structural differentiation, and when the adaptation window closes the org chart is still in review. PayPal's drift after its founding team departed is the same pattern in slow motion: product-culture erosion turning into share loss against Apple Pay and Shop Pay.

   The Enterprise AI Adoption Paradox

   Fifty Midwest enterprise CIOs gathered in one room. Zero had agents running at scale. Only twenty percent had any in production. A reasonable skeptic would read that as evidence the technology is not ready. The reasonable skeptic is wrong about the constraint. The constraint is not model quality. It is that nobody has documented how work actually gets done. The stall is buying time to restructure. It is only useful to the firms that use the stall for the reorganization rather than as permission to defer it.

   What to do

   1. →

      Classify your organization on the Ramp-to-Parker spectrum this month — count management layers below C-suite, ratio of pure managers to player-coaches, and per-employee revenue vs. AI-native peers

      This sprintThe market is assigning 25% valuation premia to AI-native org design. If your competitors are restructuring and you're not, the gap compounds quarterly.
   2. →

      Launch a process documentation initiative across your three highest-value workflows within 60 days — this is the prerequisite for any meaningful agent deployment

      This sprintEnterprise AI is bottlenecked on codified process knowledge, not model capability. Organizations that can't describe how work gets done can't automate it.
   3. →

      Stand up a token consumption dashboard and define adoption tiers within 30 days — treat internal AI usage as a product engagement metric, not an IT policy

      NowSendbird's CEO tracks token consumption like revenue because it's a leading indicator of organizational competitiveness. What you don't measure, you can't improve.
   4. →

      Model your headcount plan assuming 30-40% fewer new hires at equivalent output, redirecting savings to AI tooling — present scenarios to the board before next planning cycle

      This quarterCloudflare just gave the market permission to make this move. Boards will start asking the question whether you raise it or not.
4. 04

   Vulnerability Discovery Just Got Free — The 12-Month Security Architecture You Planned Against No Longer Exists

   The Week The AI Vuln-Economics Thesis Stopped Being Speculative

   This week's evidence makes the thesis that AI permanently changes vulnerability economics hard to treat as speculative. The results arrived from uncorrelated directions, with no shared authors and no shared codebases:

   1. Mozilla + Anthropic Mythos: 271 exploitable Firefox bugs in 60 days, 180 of them sec-high, with "almost no" false positives. The architecture is an LLM wrapped in a harness with build-tools access, and a second LLM verifying findings.
   2. Solo researcher + LLM swarm: one person, running a "homegrown swarm of LLM-powered agents," generated 20+ assigned CVEs, including remote unauthenticated out-of-bounds writes against Linux ksmbd, Docker, OpenSSL, and HAProxy.
   3. Vercel open-sourced deepsec: chains scan, investigate, revalidate, enrich, and export using Claude Opus 4.7 and GPT-5.5 across 1,000+ sandboxes with 10-20% false-positive rates. Free.

   ---

   The Offensive Side Is Worse

   Google confirmed the first AI-built zero-day exploit used by attackers in the wild, a Python script bypassing 2FA on a web administration tool, flagged through telltale LLM signatures including educational docstrings and a hallucinated CVSS score. Palisade Research clocked autonomous AI hacking success rates at 81%, up from 6% twelve months ago. That is a capability doubling roughly every three months.

   The security architecture most organizations run was designed for human-speed attackers working under human constraints. This month is when that assumption stopped holding.

   What This Means Structurally

   Detection has become commodity infrastructure, the equivalent of compute after AWS. The price of world-class vulnerability scanning moved from a $2M annual contract to roughly $200/month in API calls. Traditional AppSec vendors (Snyk, Checkmarx, Veracode) face disruption not from a better-funded startup but from open-source tools running on subscriptions customers already pay for.

   The bottleneck has shifted, not disappeared. Remediation velocity is now the binding constraint. If discovery output rises 10-50x while the engineers responsible for closing tickets do not multiply to match, every CISO inherits a vulnerability inventory that cannot be worked down with current headcount or SLAs. The disclosure-to-exploit window is compressing from weeks to hours while internal patch SLAs still assume the old timeline.

   The Internal Threat Surface

   AI agents deployed inside the business are generating non-human identities at a rate existing IAM was never designed to govern. An authenticated agent is effectively a tireless insider with perfect recall. MCP connectivity between agents and enterprise systems is almost entirely unmonitored. CTEM programs do not cover it, DLP tools do not see it, and the SOC cannot build detections against traffic it has no visibility into.

   What to do

   1. →

      Reduce mean-time-to-patch SLA by 50% for internet-facing and authentication-adjacent systems within 60 days — brief the board on the rationale this month

      NowDisclosure-to-exploit windows are compressing from weeks to hours. SLAs negotiated for human-speed attackers are now the binding constraint on defense.
   2. →

      Commission a 30-day evaluation of Vercel's deepsec against your current SAST/DAST stack — use findings to renegotiate or exit commercial scanning contracts

      NowThe price of enterprise-grade vulnerability scanning just dropped 99%. Multi-year contracts signed at the old price point are now indefensible.
   3. →

      Audit all MCP connections and AI agent data access paths — build a non-human identity inventory with permission mapping before an incident forces one

      This sprintAI agents are outrunning IAM governance. The blast radius of a compromised agent exceeds a compromised service account because agents are designed to improvise.
   4. →

      Invest in automated remediation pipeline capability — evaluate early-stage vendors or fund an internal team, budget decision within 90 days

      This quarterDiscovery just got 50x cheaper. Remediation didn't. The bottleneck moved, and it's the gap between discovery volume and fix capacity that becomes the board-reportable risk.