The Board Room

Deep dives

1. 01

   The generation bottleneck is solved — your AI engineering spend is pointed at the wrong problem

   Shopify Just Revealed Where the Real AI Engineering Gap Is

   Shopify's CTO Mikhail Parakhin — who built and shipped Sydney at Microsoft and ran Windows, Edge, Bing, and Ads — has delivered the most granular public accounting of enterprise AI transformation to date. The headline metrics are striking: near-100% daily active AI tool usage across all employees, PR merge volume growing 30% month-over-month with increasing complexity, and a December 2025 phase transition where model quality crossed a threshold making adoption self-sustaining.

   But the strategically consequential finding is this: the bottleneck has permanently shifted from code generation to review, testing, and deployment. Shopify's CI/CD pipelines are 'creaking.' No existing commercial tool meets enterprise requirements — Shopify had to build a custom PR review system using the most expensive frontier models available (GPT 5.4 Pro, Deep Think from Gemini). The entire $15-20B AI coding tool market is optimized for the problem that's already solved.

   The company that builds enterprise-grade AI code review — not at Copilot's level but at frontier-reasoning level — captures the next layer of developer productivity value.

   Cloudflare Proves AI Review Works at Production Scale

   While Shopify describes the gap, Cloudflare has started filling it. Their AI code review system processed 131,246 reviews in month one as a mandatory pipeline gate across all engineering. Key metrics: $1.19 per review, 3 minutes 39 seconds median latency, and a 0.6% override rate — meaning engineers almost never override the AI reviewer. They built custom using seven specialized agents with circuit breakers, model failback chains, and an 85.7% cache hit rate. The signal: the most sophisticated buyers are building, not buying — commercial tools aren't meeting enterprise needs yet.

   Shopify's three proprietary systems reinforce this pattern. Tangle (open-source ML experimentation with content-addressed caching), Tangent (auto-research loops so effective a PM is the top user, delivering 5x search throughput improvements expert teams hadn't found), and SimGym (customer simulation at 0.7 correlation with real behavior) — together these convert Shopify's data assets into compounding competitive advantage that no vendor can replicate.

   ---

   The Token Cost Explosion You're Not Tracking

   Simultaneously, the economics of AI compute are fragmenting in ways most finance teams haven't modeled. Token pricing has splintered into 8+ distinct SKUs billed at wildly different rates. Reasoning tokens inflate actual costs 10-15x above what visible output suggests — and providers haven't standardized how they report or bill for these categories. This is an information asymmetry that advantages sellers.

   The subsidy era is ending in parallel. GitHub shifted to token-based billing and paused new signups. Anthropic is testing $100/month Claude Code pricing. Analysis shows identical AI services priced at 185x differences across providers. Cloudflare consumed 120 billion tokens in one month of AI code review alone. As you layer AI across review, security scanning, alert triage, and code generation governance, inference costs compound — and the CFO conversation about 'AI infrastructure costs' arrives whether you initiate it or not.

   If your product charges customers $X per AI interaction but your underlying cost varies 2-15x depending on reasoning tokens, you have a pricing model vulnerability that worsens with scale.

   The Cross-Source Pattern

   Multiple sources converge on one conclusion: the SaaS P&L model assumed 70-80%+ gross margins with near-zero marginal costs. AI destroys this assumption. Every API call, every vector DB query, every model routing decision is a variable cost that scales with usage. Companies burying these in generic cloud infrastructure line items will discover the problem when growth isn't translating to margin expansion. The companies that instrument AI COGS visibility at the board level now — isolating inference, model routing, vector DB, and embedding costs per customer — will navigate this repricing. Those that don't will face a margin crisis they can't diagnose.

   What to do

   1. →

      Audit your AI token spend by type (input, output, reasoning, cached) across all production workloads this sprint — build a dashboard showing cost-per-interaction trending

      NowReasoning tokens alone can inflate costs 10-15x; without visibility, you're flying blind on margin as you scale
   2. →

      Measure your code review latency (generation-to-merge time) and benchmark against Cloudflare's 3m39s median and 0.6% override rate by end of Q3

      This sprintShopify confirms review is the binding constraint — companies that solve it capture the next productivity layer
   3. →

      Build a model abstraction layer into production AI systems — no workload should be hard-coupled to a specific model version's billing structure or behavioral quirks

      This quarterOpus 4.7 migration breaks prompting patterns, thinking configs, and tool-calling behavior — abstraction eliminates recurring migration tax
   4. →

      Model your AI engineering budget at 3-5x current Copilot/Claude costs and present to CFO as a scenario

      This sprintGitHub subsidy end and Anthropic's $100/month test signal normalized pricing is materially higher than today's rates
2. 02

   Three structural breaks in cybersecurity this week — your vulnerability and incident response models both failed

   NIST Concedes Defeat on CVE Enrichment

   As of April 15, NIST will only enrich CVEs appearing in CISA's KEV catalog, affecting federal software, or qualifying as critical under Executive Order 14028. This isn't a temporary resource constraint — it's an institutional acknowledgment that CVE volume has permanently outstripped the public-good model. If your vulnerability management program, SLAs, or board-level risk metrics depend on NVD severity scores, your security posture visibility is degrading right now.

   The second-order effect is a widening gap between organizations that can afford commercial vulnerability intelligence (VulnDB, Snyk, Qualys feeds) and those that cannot — effectively creating a two-tier security ecosystem where resource-constrained organizations lose the ability to prioritize. For security vendors, this is a once-in-a-decade market-creation event.

   ---

   Ransomware Terrorism Designation Gains Congressional Traction

   Former FBI Cyber Deputy Director Cynthia Kaiser testified before the Homeland Security Committee with a specific legal framework: terrorism designation for hospital ransomware. The data supporting urgency: FBI figures show healthcare ransomware incidents nearly doubled from 238 to 460 between 2024 and 2025. A University of Minnesota study linked dozens of Medicare patient deaths to these attacks. The committee chairman responded that 'no penalties are too severe.'

   The second-order effect matters more than the first: if ransomware groups targeting hospitals are designated terrorists, paying ransom could constitute material support for terrorism. This would eliminate ransom payment as an incident response option, forcing a wholesale shift toward resilience and recovery. Treasury has already proposed extending terrorism risk insurance (TRIA) to cover cyber losses. Connect the dots: mandatory security standards, TRIA-backed insurance coverage, and potentially homicide prosecution within 12-24 months.

   A former senior FBI cyber official is building the evidentiary case for ransomware-as-terrorism — and she's now at a company commissioning updated mortality data to support it. This isn't advocacy; it's prosecution prep.

   Your IR Vendor May Be Working for the Attacker

   The Martino guilty plea at DigitalMint isn't an isolated scandal — it's a structural vulnerability in the incident response ecosystem. A ransomware negotiator used inside knowledge of victim insurance limits, negotiating posture, and willingness to pay to help BlackCat/ALPHV affiliates extort the companies that hired him. He then conspired with other IR professionals to deploy ransomware against additional firms. $10 million in seized assets confirms this was sophisticated and profitable.

   Every enterprise with an IR retainer is giving vendors the exact intelligence that maximizes ransom extraction. Expect CISO purchasing committees to demand compartmentalized access, background verification, and audit trails for IR engagements. This is a new product category — IR governance and vendor trust — being born right now.

   ---

   AI Collapses the Patch Window to Near-Zero

   The vulnerability lifecycle is fundamentally broken across multiple dimensions simultaneously. AI discovers bugs at machine speed (Anthropic's Mythos finding 271 Firefox zero-days, as previously reported). LLMs now weaponize disclosed vulnerabilities within minutes. But enterprises still patch in 12+ days. Meanwhile, AI agents themselves are introducing novel attack surfaces: Azure SRE Agent's multi-tenant authentication flaw exposed live command streams and credentials to any Entra ID account holder. Google's Antigravity IDE sandbox escape via prompt injection — where the agent executed a shell command before Secure Mode could evaluate it — represents an entirely new attack class with no established defense model.

   The protobuf.js vulnerability (CVSS 9.4) illustrates the supply chain dimension: it silently affects any system touching Firebase, gRPC, or Google Cloud SDKs via transitive dependency. Most teams won't know they're exposed without a deep SCA scan. Combined with research showing 72% of 6,121 public Perforce servers allow unauthenticated read access, the foundational trust assumptions in your security architecture are being systematically invalidated.

   What to do

   1. →

      Audit your vulnerability management pipeline for NVD enrichment dependency this week — identify every tool and SLA that assumes NVD provides timely enrichment for non-KEV CVEs

      NowNIST policy shifted April 15; your risk dashboards may already be silently degrading
   2. →

      Mandate an emergency SCA scan for protobuf.js exposure across all production services — flag any dependency on versions prior to 8.0.1 or 7.5.5, including Firebase and gRPC transitive exposure

      NowCVSS 9.4 RCE with trivially exploitable pattern; most teams won't know they're exposed
   3. →

      Audit your incident response vendor relationships — specifically how insurance data, negotiation posture, and willingness-to-pay intelligence is compartmentalized

      This sprintMartino case proves IR vendors have the exact intelligence needed to maximize extortion
   4. →

      Convene a cross-functional review of your ransomware response playbook, specifically the ransom payment decision tree, with outside counsel on evolving terrorism designation legal exposure

      This quarterIf ransomware groups get terrorism designation, ransom payments become material support — your IR playbook needs to reflect this before it happens
3. 03

   The model layer is commoditizing on a timeline measured in months — where value goes next

   Open-Weight Parity Is No Longer Theoretical

   The evidence converged this week from multiple angles. Open-weight Kimi K2.6 matches or exceeds Claude Opus 4.6 on four of six head-to-head benchmarks including SWE-bench Pro (58.6 vs 53.4) at one-fifth the cost. Apple — with $100B+ in annual R&D — chose to power Siri's major overhaul with Gemini models rather than build its own, a capitulation that signals even the world's richest company sees model development as a losing bet. Google's Gemma 4 ships fundamentally different architectures for edge vs. server, breaking the 'one model scaled up or down' paradigm entirely. The model layer is becoming infrastructure, not differentiation.

   Simultaneously, Anthropic is showing execution cracks: removing Claude Code from Pro plans while Opus 4.7 users report reliability degradation. Reddit discussions show users doing explicit cost-benefit analysis and choosing open-weight alternatives at $19/month against $20/month Claude Pro. The timing couldn't be worse for a company raising at frontier valuations.

   If Apple — with $100B+ R&D spend — concluded that building competitive foundational models isn't worth the investment, that signal should penetrate every boardroom.

   Google's Silicon Split Is the Infrastructure Signal

   Google bifurcating its eighth-generation TPU into dedicated training (8t) and inference (8i) chips — the first time in TPU history — is a structural market signal. The 8i: 288GB of high-bandwidth memory, 384MB of on-chip SRAM, and a new Boardfly network topology delivering up to 5x latency reduction. Google is saying, in silicon, that inference workloads justify dedicated chip design, dedicated supply chains (Broadcom for training, MediaTek for inference), and dedicated optimization.

   Gemma 4's MoE economics reinforce this: 6.25% activation ratio means the 26B model stores 25.2B parameters but activates only 3.8B per token, delivering 70B-class reasoning at 8B-class inference cost. Any company still running dense 70B models for production inference is paying a 5-8x premium that MoE adopters will exploit. Caveat: Gemma 4 has a critical Blackwell GPU dependency — on pre-Blackwell hardware, throughput drops 14x to ~9 tokens/second.

   ---

   a16z Telegraphs What Comes After RAG

   When Andreessen Horowitz publishes that continual learning is 'some of the most important work happening in AI right now,' that's a public investment thesis, not a research survey. The core argument: today's LLMs are frozen at deployment, and everything we do post-training — RAG, prompt engineering, context management — is compensating for the fact that models can't actually learn from experience. a16z frames these as bridge technologies with a 2-3 year shelf life.

   The practical implications are staggering. An 8B model with targeted continual-learning modules matching 109B performance on targeted tasks represents a potential order-of-magnitude reduction in inference costs for domain-specific applications. The Ilya Sutskever framing — pre-training 'overshot the target' by compressing everything at once — suggests the field's most important researcher is reorienting around this problem.

   For your infrastructure strategy: if you've invested heavily in RAG pipelines, vector databases, and context-management tooling, a16z views these as bridge technologies. The 3-year depreciation schedule may be generous. But the companies that build on parametric learning will offer something structurally different: AI that develops genuine expertise encoded in weights, not by looking things up. The strategic play is building competitive advantage in layers valuable under either scenario — workflow integration, proprietary data flywheels, and domain-specific agent orchestration.

   What to do

   1. →

      Run a 30-day parallel evaluation of Kimi K2.6 against your current Anthropic/OpenAI workloads on your highest-volume, most cost-sensitive pipelines

      This sprintOpen-weight delivers 85% capability at 81% cost reduction — failing to test this is leaving money on the table
   2. →

      Build an intelligent model routing layer that dispatches tasks to different models and effort levels based on complexity, latency, and cost this quarter

      This quarterThe 5x cost spread between open and closed models on equivalent tasks means misrouting is pure margin destruction
   3. →

      Commission a strategic review of your RAG infrastructure investments with explicit depreciation scenarios under a16z's continual learning thesis

      This quartera16z signaling capital flow means startups will ship production-grade continual learning within 18 months; your RAG stack may have a shorter useful life than planned
   4. →

      Track Google's Blackwell TPU dependency and model your inference cost curve assuming dedicated training/inference silicon becomes standard within 18 months

      WatchGoogle's silicon split signals inference economics will shift dramatically; early movers on specialized silicon capture structural cost advantages