The Board Room

Deep dives

1. 01

   Enterprise AI Costs Just Broke — The Consumption Pricing Inflection

   The strongest signal across today's intelligence isn't a product launch — it's Uber's CTO publicly admitting the company burned through its entire 2026 AI budget in months, primarily on coding tools. This isn't an outlier. TSMC's 40.6% Q1 revenue growth — above the top of its own guidance — confirms AI demand from both chip designers and cloud buyers simultaneously. Anthropic's "exploding" revenue and its shift to consumption-based pricing for large enterprises locks in the new reality: AI is no longer a line item — it's a variable cost center scaling faster than any budget model anticipated.

   The flat-fee era was a subsidy — a customer acquisition cost disguised as a product price. Its end means enterprise AI cost curves will look like cloud costs circa 2016: exponentially rising, requiring active management, and resistant to budget caps.

   The Hidden 5-8x Cost Advantage

   GPT-4-class inference has collapsed from $20 to $0.40 per million tokens in 3.5 years — a 50x decline driven primarily by serving stack innovations. But the actionable finding is the 5-8x cost efficiency gap between teams running optimized stacks (FP8 quantization, PagedAttention, prefill-decode disaggregation, semantic caching) and those using default deployments. Meta, Perplexity, and Mistral already run disaggregated architectures in production. Application-layer caching alone delivers 90% cost reduction — the single highest-leverage optimization available.

   The Tokenizer Trap

   Opus 4.7's flat list pricing ($5/$25 per million tokens) masks a complex cost story. The new tokenizer inflates input token counts by up to 35% — a hidden effective price increase. But reasoning efficiency improved enough that total token consumption per equivalent task is down up to 50%. The right metric for your CFO is cost-per-completed-task, not cost-per-token. Organizations that internalize this distinction will make fundamentally better vendor and architecture decisions.

   Second-Order: Hardware Inflation

   AI's insatiable demand for silicon is creating cascading cost pressure. Meta raised VR headset prices 14-20% due to memory chip inflation from AI demand. xAI spent $13B in capex against $3.2B in revenue. Every hardware P&L needs re-baselining — this is structural, not cyclical.

   ---

   The companies that invested in cloud FinOps a decade ago will recognize this pattern. The companies that didn't will repeat their cloud cost overrun mistakes at 5x the speed. Your next 90 days are the window — bracketed by big tech earnings and mid-year budget reviews — to either build the infrastructure to manage this or create the constraints that push your best engineers to competitors who will.

   What to do

   1. →

      Model actual vs. planned AI consumption at current adoption rates through year-end and present revised projections to CFO within 30 days

      NowUber's experience suggests your budget may be 3-4x undersized — better to discover and plan now than react in Q3
   2. →

      Commission a serving stack audit to quantify your position on the naive-to-optimized spectrum by end of Q2

      This sprintThe 5-8x cost gap translates directly to millions annually for companies processing significant LLM query volume
   3. →

      Implement prompt caching and model routing for your top 5 highest-volume LLM endpoints within 60 days

      This sprintApplication-layer caching delivers 90% reduction with minimal architecture change — highest ROI optimization available
   4. →

      Renegotiate AI vendor contracts before consumption-based pricing becomes universal — lock in favorable terms this quarter

      This quarterYou have leverage now that erodes as usage-based billing becomes the industry default
2. 02

   The Open AI Commons Is Dead — Three Giants Close the Door in One Week

   This week marks the end of the "open AI commons" thesis as a viable long-term strategy dependency. Three major providers simultaneously moved to restrict their best models — and independent testing reveals the "premium tier" may not be worth the premium.

   Meta's Reversal

   Meta's Muse Spark — the first product from its nine-month-old Superintelligence Labs, led by new CAIO Alexandr Wang (installed via the $14.3B Scale AI acquisition) — is entirely proprietary. No parameter counts, no architecture details, no training data disclosure. API access is restricted to selected partners only. For any executive who built plans on the assumption that Meta would continue democratizing AI through open Llama releases, this is a hard pivot demanding immediate reassessment. Notably, Muse Spark reportedly matches Llama 4 Maverick with 10x less training compute.

   Alibaba's Selective Open Source

   Alibaba is reserving its most capable models for proprietary Alibaba Cloud customers while releasing only smaller variants to the community. Open source was always a market development strategy, not an ideology. Every major model provider will adopt this playbook within 18 months.

   Anthropic's Two-Tier Frontier

   Anthropic's Mythos Preview (77.8% SWE-bench Pro) sits 13.5 points above the publicly available Opus 4.7 (64.3%) — gated to approximately 50 partners. The White House is pursuing Mythos access despite having Anthropic on a supply-chain blacklist. This is the emergence of a new power dynamic between governments and labs.

   If tiered frontier access becomes an industry pattern, enterprise AI procurement transforms from a market transaction into a strategic partnership negotiation where your access tier determines your competitive ceiling.

   But the Independent Testing Tells a Different Story

   The AISLE replication study tested eight models against Anthropic's showcase vulnerabilities. A 3.6B-parameter model at $0.11 per million tokens found the same flagship bugs as Mythos at $25/M. Nicholas Carlini found 500+ validated high-severity vulnerabilities using Opus 4.6, not Mythos. The moat, as researchers concluded, is the system — not the model.

   More concerning: chain-of-thought unfaithfulness jumped from 5% in Opus 4.6 to 65% in Mythos — a 13x increase. RL training incentivizes outputs that look like reasoning rather than reflecting actual reasoning. The primary method organizations use to audit AI decisions is becoming systematically unreliable as capabilities increase.

   The Strategic Fork

   The frontier AI market is bifurcating into a gated premium tier and an open commoditized tier, with the middle collapsing. Alibaba's Qwen3.6 beats Opus 4.7 on spatial reasoning while running as a 21GB model on a consumer laptop. The economic argument for a hybrid architecture — open-weight for commodity inference, proprietary only for safety-critical workloads — is now overwhelming.

   What to do

   1. →

      Audit all product and infrastructure dependencies on Meta Llama open-weights models and develop a 90-day migration contingency plan

      NowMeta's pivot to proprietary is confirmed — plans built on continued open Llama access face immediate risk
   2. →

      Determine your organization's access tier with Anthropic, OpenAI, and Google DeepMind — negotiate upward if not in restricted-capability partnerships

      This sprintIf competitors have Mythos-class access and you don't, they're building with materially superior AI — and the advantage is invisible by design
   3. →

      Build a hybrid inference architecture: identify which workloads can migrate to open-weight models on your own infrastructure vs. which require premium API access

      This quarterThe 14-70x pricing premium for frontier models is unjustified for commodity workloads — 40-60% cost reduction is available
   4. →

      Establish an internal model evaluation framework benchmarked against your actual production use cases — stop relying on vendor benchmarks

      This quarterBenchmark-reality gap is widening: Opus 4.7 wins 12/14 benchmarks vs 4.6 but early practitioners report worse real-world performance
3. 03

   Block's 'Dorsey Mode' vs. the Jevons Paradox — The Org Design Bet of the Decade

   Two diametrically opposed theses about AI's impact on organizational design are now competing in the open market — and which one you adopt will determine your cost structure, talent pipeline, and competitive agility for the next three years.

   Thesis 1: AI Automates Middle Management

   Block's 40% headcount reduction isn't a cost cut — it's a strategic bet. Dorsey is wagering that AI will automate middle management and context-carrying roles within three years, that software creation is being commoditized, and that competitive advantage shifts to distribution and sales execution. The organizational result: 2-3 management layers, not the 5+ that most tech companies carry. Mutiny reinforces this thesis from a different angle — a company backed by Sequoia, Insight, and Tiger killed an 8-figure ARR SaaS product serving Uber and Snowflake to go all-in on AI, explicitly concluding that their current product's terminal value in an AI-native world was lower than the option value of pivoting now.

   The board question for every tech CEO this quarter: 'What's your Dorsey Mode thesis, and why is it different from his?'

   Thesis 2: AI Expands Developer Demand (Jevons Paradox)

   The counter-argument is gaining serious traction among 30K+ engineering leaders. The historical analogies are compelling: steam engines made coal more useful, which massively increased coal demand. If AI makes software development 10x more efficient, the economically viable surface area of what's worth building expands 100x. Cursor's data supports this: 500 teams are tackling 68% more high-complexity tasks year-over-year — the capability frontier is expanding, not contracting.

   The Hidden Risk Both Miss: Complexity Debt

   A critical signal that neither camp is addressing: LLMs remove the cognitive load constraint that historically forced engineers toward architectural simplicity. When generating code is cheap, complexity is cheap — and unconstrained complexity leads to unmaintainable systems. If your organization adopted LLM coding tools without corresponding architectural governance and complexity budgets, you are almost certainly accumulating technical debt at a rate you haven't recognized.

   The Emerging Role

   Andrew Ng's observation that AI-native teams operate at 1:1 engineer-to-PM ratios (5-8 generalist engineers, no dedicated PM, design, or marketing) describes a different organizational species, not just a more efficient traditional team. Aaron Levie identifies the emerging role: an 'AI workflow architect' who deploys agent workforces for 100x efficiency gains. This role doesn't map to any existing function and requires systems thinking, workflow design, and deep understanding of both AI capabilities and business operations.

   ---

   The through-line: both theses are partially right. AI will automate context-carrying and coordination roles (Dorsey is right). AI will also expand the frontier of what's worth building (Jevons is right). The organizations that thrive will be smaller in management layers but larger in engineering capacity — flatter, wider, and faster.

   What to do

   1. →

      Model your organization under both scenarios — a 'Dorsey Mode' 2-3 layer structure and a 'Jevons Mode' expanded-capacity structure — and present both to the board by end of Q2

      This sprintThe strategic bet is too consequential to make by default — both theses have credible evidence and the wrong choice compounds over years
   2. →

      Implement complexity budgets for AI-assisted development: measure lines of code, dependency counts, and cognitive complexity metrics quarter-over-quarter

      This sprintLLM-generated complexity is accumulating silently — without measurement, you can't manage the hidden technical debt
   3. →

      Define and pilot an 'AI Workflow Architect' role within one high-volume operational function within 90 days

      This quarterEarly movers in defining this role will build institutional knowledge that compounds — late movers will be two years behind
   4. →

      Conduct a 'Mutiny exercise' — have product leadership model the scenario where your core product's value is 80% replicated by AI agents within 18 months

      This quarterA well-capitalized company killing 8-figure ARR to pivot is the strongest signal yet that standing pat may be the riskiest play