The Board Room

Deep dives

1. 01

   Google's Two Bombs: 2029 Quantum Deadline and 6x Inference Compression Force Simultaneous Planning Resets

   Google dropped two infrastructure signals this week that, taken together, invalidate the cost assumptions and the security assumptions underlying most enterprise AI strategies. Both demand action this quarter — not because the sky is falling, but because the organizations that move first will capture structural advantages that compound for years.

   ---

   The Quantum Clock Just Lost Six Years

   Google moved its internal post-quantum cryptography migration deadline from 2035 to 2029 — six years ahead of NIST's federal baseline. This isn't a research position; they're already shipping PQC in Android 17 beta for developer signing keys. When a company with the world's most advanced quantum computing program begins hardening production systems, that's the strongest possible signal that the threat timeline has compressed.

   When Google says 'harvest now, decrypt later' attacks are already active, they're telling you the threat window isn't 2029-forward — it's now. Every piece of encrypted data with a secrecy shelf-life beyond 5 years is potentially compromised the moment a capable quantum machine comes online.

   The White House is actively considering pulling the federal deadline to 2030 or earlier. When that happens, the regulatory cascade is predictable: FedRAMP, CMMC, and sector-specific regulators will all follow. Companies that have already begun migration gain procurement advantages; those that haven't face compressed, expensive timelines.

   The strategic calculus is a classic first-mover problem: organizations that begin crypto-agility engineering now can deploy NIST algorithms as a configuration change. Those that wait until 2028 face a talent war for PQC expertise and a vendor scramble that will make Y2K look orderly.

   ---

   TurboQuant Breaks the AI Cost Curve Through Software, Not Hardware

   Google Research released TurboQuant: 6x KV cache memory reduction, 8x attention speedup, zero accuracy degradation — benchmarked on production H100 hardware. Wall Street reacted immediately: Micron and Western Digital dropped 3-5%. But the stock move understates the structural shift.

   If inference memory requirements drop 6x across the industry, the competitive moat shifts from 'who has the biggest GPU fleet' to 'who has the most efficient inference stack.' This benefits algorithmic innovators (Google, Meta) and threatens companies whose strategy depends on infrastructure scale as a barrier to entry.

   Combined with Apple's newly secured Gemini distillation rights — running frontier models on-device without internet — the inference economics are converging on a world where capable AI runs at the edge, not in the cloud. For any company in healthcare, financial services, defense, or privacy-sensitive domains, on-device inference eliminates the most significant objection to AI adoption: sending data to someone else's cloud.

   The companies that win the next cycle won't have the biggest GPU fleet — they'll have the most efficient inference stack. Hardware scale as a moat is eroding through software innovation.

   What to do

   1. →

      Commission a cryptographic agility audit across all products, infrastructure, and third-party integrations — inventory every RSA/ECC dependency

      NowGoogle's 2029 deadline means your migration must start this fiscal year to avoid a compressed, chaotic scramble in 2028-2029
   2. →

      Revise 2026-2028 AI infrastructure capex projections downward by 40-60% for inference workloads, incorporating TurboQuant-class compression

      This sprintCurrent projections likely overstate RAM and compute requirements; the paper is being presented at ICLR in April and every major lab will integrate by Q3
   3. →

      Launch an edge inference feasibility study for your most privacy-sensitive or latency-critical AI use cases

      This quarterTurboQuant + Apple distillation signals the inference cost curve will favor edge deployment within 18 months — first movers in regulated industries capture disproportionate trust advantage
   4. →

      Negotiate flexibility clauses into any GPU/compute procurement contracts signed this year

      This sprintCompute scarcity assumptions underlying current contracts may not hold as algorithmic efficiency gains propagate
2. 02

   The AI Strategy Paradox: Sub-1% Reasoning + $50M Frontier Models = Every Assumption Needs Stress-Testing

   Two data points released this week appear contradictory but together paint the clearest strategic picture of where AI actually stands — and it's not where most boardrooms think.

   ---

   The Reasoning Ceiling Is Real and Severe

   ARC-AGI-3 tested frontier models on 135 interactive mini-games that every human solves on the first try — not PhD math, but scenarios requiring rule discovery, goal formation, and strategy. The results are stark:

   Model	Score
   Humans	100%
   Gemini Pro 3.1	0.37%
   GPT-5.4 High	0.26%
   Opus 4.6	0.25%
   Grok	0%

   This isn't a marginal miss — it's a categorical failure revealing that current architectures lack something fundamental about human reasoning. Separately, AI-generated code shows a 48% hallucination rate (o4-mini), and a pentest of a 100% Claude Opus 4.6-generated application found critical local file inclusion, insecure direct object references leaking password hashes, and outdated dependencies with known CVEs.

   Current AI is extraordinarily capable at pattern matching and generation, but fundamentally limited at novel reasoning. Any product roadmap assuming near-term autonomous agents is building on a foundation that doesn't exist yet.

   ---

   The Cost Floor Is Collapsing Anyway

   While capabilities plateau, access to frontier-class AI is democratizing at unprecedented speed. Xiaomi's Hunter Alpha — a trillion-parameter model with a million-token context window — appeared anonymously on OpenRouter, processed 160 billion+ tokens, and was widely assumed to be DeepSeek v4. When Xiaomi was revealed as the creator, their stock jumped 5.8%. Users could not tell the difference between a smartphone company's model and a frontier lab's.

   Nvidia-backed Reflection is raising $2.5B at $25B, positioning as the 'DeepSeek of the West' — reframing open-source AI from a cost play to a sovereignty play. Verkor AI demonstrated an AI agent designing a complete Linux-capable 1.5 GHz RISC-V CPU from concept to tape-out in 12 hours (180x faster than a human team). The recursive loop has closed: AI designs chips, better chips train better AI.

   The 'monetizable spread' between open and closed-source AI is declining faster than the capability spread. The premium customers will pay is compressing even where gaps persist. Fast-followers replicate frontier results at a fraction of cost because most R&D spending goes to exploration that subsequent entrants learn from cheaply.

   ---

   The Strategic Synthesis

   Your AI strategy is likely simultaneously too optimistic on capabilities and too pessimistic on access. The correction needed is precise: invest aggressively in AI for pattern-matching, summarization, and structured automation (where it excels), while maintaining deep skepticism about reasoning-dependent autonomy. Build for human-in-the-loop as the primary architecture through 2028, with autonomous as progressive enhancement. Your moat isn't model access — it's proprietary data, domain expertise, and workflow depth.

   What to do

   1. →

      Audit every AI product roadmap against genuine reasoning requirements versus pattern-matching tasks — flag any feature assuming autonomous novel reasoning for deferral or redesign

      This sprintARC-AGI-3 proves autonomous reasoning is years away, not quarters; overcommitting resources to autonomy creates credibility risk when it inevitably underdelivers
   2. →

      Stress-test every AI vendor relationship and investment against a world of $50M frontier models — model how open-source parity affects pricing leverage over 8, 16, and 24-month horizons

      This quarterThe monetizable spread is collapsing; vendor contracts signed under scarcity assumptions become liabilities
   3. →

      Launch an aggressive proprietary data acquisition and curation initiative — inventory every source of unique, domain-specific data your organization generates or could acquire

      This quarterWhen model access is universal, proprietary data is the only durable moat; organizations that haven't begun systematic collection are giving away competitive advantage
   4. →

      Implement mandatory automated security scanning (SAST, DAST, dependency checks) for all AI-generated code before production merge

      Now48% hallucination rates and demonstrated critical vulnerabilities in AI-generated code create both security risk and emerging product liability exposure
3. 03

   Enterprise AI Hits the Kill-or-Scale Wall: Novo Nordisk's $100M/Week Framework vs. 68% Pilot Purgatory

   The enterprise AI market is bifurcating sharply between organizations that have found quantifiable ROI and those still running pilots. The data from this week paints a picture of an industry exiting the experimentation phase whether companies are ready or not.

   ---

   The ROI Benchmark That Changes the Conversation

   Novo Nordisk's CDO has done what most enterprise AI deployments cannot: quantify AI agent value in board-level terms. Each week of clinical trial acceleration equals tens to hundreds of millions in peak revenue. Their AI agents now handle data analysis previously outsourced to several hundred contractors and detected that three internal departments were redundantly overseeing regulatory document drafting.

   But here's the critical counterpoint: Novo also deliberately killed its own AI tool (Found Data, powered by Anthropic's Claude) because it was expensive and didn't yield noticeable advances. CDO Stephanie Bova's standard is brutal: 'If I can do it better and cheaper and more reliably in Excel, I'm going to tell you to stay in Excel.'

   Their architecture is also instructive: Celonis as the orchestration layer, routing tasks to optimal models across Anthropic, OpenAI, and others. The orchestration layer — not the model — is where enterprise value consolidates. If your platform strategy assumes single-vendor AI lock-in, Novo's multi-model approach directly challenges that assumption.

   ---

   The 68% Problem

   Across the S&P 500, over 1,000 AI partnerships have been announced — growing 23% in two years. But 68% remain at pilot, integration, or co-marketing stage. Only 12% are actual production vendor/client relationships. This is the most dangerous phase: buyers scatter investment across experiments without building real capability, while sellers burn runway on pilots that never convert.

   The consolidation signal is equally clear: Nvidia, Microsoft, and Amazon are emerging as the three dominant AI partnership hubs. Enterprise lock-in decisions are being made now, during the pilot phase, often without deliberate strategic intent.

   ---

   The FDE Bottleneck Reveals Product Immaturity

   Field Development Engineer postings exploded 10x in 2025, but only ~10% of engineers want the role. Every FDE req is an admission that the product needs a human translator. Companies are compounding the problem by misrepresenting FDE roles as engineering positions — engineers accept, discover they're doing sales support, and churn within weeks, poisoning the talent pool further.

   The structural read: if your enterprise GTM depends on FDE-type roles scaling linearly with revenue, you're building a services business with a software valuation. The solution isn't hiring more FDEs — it's investing in product self-service maturity to eliminate the dependency.

   The era of 'let's experiment with AI' is over. Your enterprise customers will demand the kind of ROI clarity Novo's CDO articulated: don't show me a business case — show me time-to-value that's so obvious it doesn't need one.

   What to do

   1. →

      Categorize every AI partnership and pilot as 'convert to production within 90 days' or 'kill' — eliminate pilot purgatory

      This sprint68% pilot-stage rate means most organizations are scattering investment without building capability; ruthless conversion-or-kill discipline separates winners from tire-kickers
   2. →

      Develop a 'time-to-revenue acceleration' value framework for your AI products modeled on Novo's weeks-saved-equals-millions approach

      This quarterEnterprise buyers entering ROI discipline phase will demand this level of specificity; vague productivity claims no longer close deals
   3. →

      Audit your deployment model: calculate FDE/SE headcount-to-revenue ratio and model whether it scales to 3-year targets — if not, redirect investment to self-service tooling

      This quarter10x FDE demand against 10% interest rate means the talent pool is structurally insufficient; product self-service is the only scalable solution
   4. →

      Position AI agent products against the outsourced contractor budget line item, not internal headcount

      This sprintNovo's displacement sequence — contractors first, then internal redundancy discovery — is the path of least organizational resistance and fastest ROI realization