ClarityBeta

Product & Strategy

The Product Desk

7 min read · 1,386 words · 3 moves to watch
The Signal

GPU prices are up 50% and causing product cancellations

Meta paid $2B for Manus's agent harness, not its model. The message across all three signals is identical: stop paying premium for raw model power and start investing in the orchestration, reliability, and collaborative UX layers where users and acquirers actually see value.

In Play

  1. Users Want Collaborative AI, Not Automation — At Scale

    Canva's 265M MAU overwhelmingly prefer collaborative AI with control over 'make it for me' automation. Anthropic's 81K-person survey confirms unreliability is the #1 user concern — ahead of job loss and autonomy. Replit's 'Deploy→Publish' rename drove 10% conversion lift. The market has shifted from 'does AI work?' to 'can I trust it?'

    Ask Clarity

  2. GPU Costs +50% and the $2B Harness Signal Reset AI Economics

    GPU prices surged ~50%, causing service outages and actual product cancellations. Meta paid ~$2B for Manus's agent harness — not its model — confirming orchestration is where value concentrates. GRPO+RULER now lets you fine-tune small models to beat GPT/Claude on specific tasks with zero labeled data. The cost-capability frontier demands hybrid architectures now.

    Ask Clarity

  3. Enterprise AI Security Is Now a Revenue Gate

    Shadow AI is the #1 CISO concern for 2026. Microsoft Copilot surfacing board decks to interns via stale SharePoint ACLs is described as 'a self-inflicted breach.' AI coding tools are hallucinating package names that attackers already squat. Wharton proved persuasion techniques 2x+ LLM guardrail bypass rates. Prompt injection is now called 'the new SSRF.'

    Ask Clarity

  4. LLMs Are Sabotaging Your Distribution — And Nobody's Tracking It

    Python's uv has 30% adoption despite near-universal developer admiration — because LLMs trained on stale data recommend pip. This is a new distribution failure mode: AI intermediaries recommending legacy patterns. Meanwhile, Canva is embedding as the 'visual execution layer' inside ChatGPT, Claude, and Gemini. AI-native distribution is the moat now; model quality is not.

    Ask Clarity

  5. Global AI Stack Bifurcation: CUDA vs. Huawei Splits Your TAM

    DeepSeek V4 is adapting to Huawei Ascend silicon. Jensen Huang says 50% of AI developers are in China. European governments are actively migrating away from US tech infrastructure. The CUDA-dependent AI stack is fracturing into regional ecosystems, and the window to architect for portability is closing.

    Ask Clarity

Deep Dives

The Collaboration Imperative: 265M Users and 81K Respondents Just Told You to Stop Building Magic Buttons

The Largest User Research Dataset You'll Get For Free

Two data sources released this week collectively represent the most comprehensive signal on what users actually want from AI products — and the answer contradicts the industry's default playbook. Canva's 265M+ monthly users overwhelmingly prefer collaborative AI with control over full automation. They describe edits in subjective language — 'make it feel more premium,' 'something warmer' — and expect the AI to interpret, not replace them. Separately, Anthropic surveyed 81,000 people across 159 countries and found unreliability is the #1 concern — ahead of job displacement, loss of autonomy, and cognitive atrophy.

If your Q3 roadmap is stacked with new AI capabilities but you haven't invested proportionally in accuracy, confidence scoring, and graceful degradation, you're optimizing for the wrong variable.

The Edit-Sequence Moat You're Probably Sitting On

Canva's technical differentiation isn't which foundation model they use — it's that they trained their model on the actual sequence of edits users made, not just finished outputs. They also deliberately 'perturb' designs (breaking spacing, hierarchy, alignment) to train error recognition. The result: emergent capabilities appeared — like converting ASCII diagrams to designs — that they never explicitly trained for. Canva's CPO says plainly: 'Most people genuinely don't care what model is running under the hood — they're not shopping for AI, they're trying to get something done.'

If your users edit documents, refine queries, iterate on configurations, or revise plans inside your product, you're sitting on process data that could power a proprietary AI capability. Most PMs log events for analytics; few structure them for model training. The gap between those two approaches is worth millions in defensibility.

Cognitive Atrophy Is a Named User Fear — Design Around It

Anthropic's survey surfaced cognitive atrophy as a real concern: users worry AI is eroding their skills. Replit's parallel finding reinforces this — renaming 'Deploy' to 'Publish' drove a 10% lift in published applications, their single biggest growth lever, because the original term intimidated the non-technical users their platform targets. The takeaway: products that position as 'AI that makes you better' (showing reasoning, teaching patterns, maintaining user decision-making) will outperform 'AI that does it for you' on long-term retention.

The 'Last Mile' Is the Real Product Opportunity

Canva is embedding itself as the 'visual execution layer' across ChatGPT, Claude, Copilot, and Gemini simultaneously — positioning not as a competitor to AI assistants but as the indispensable step where ideas become publish-ready output. The framing: chatbots are 'great for thinking, a dead end for doing.' Does this pattern exist in your domain? In legal, AI drafts contracts but can't format for filing. In data, AI generates analysis but can't produce board-ready visuals. These last-mile gaps are where massive product opportunities live.

What to do

  1. Audit your top 5 AI features for 'automation bias' this sprint — any feature that removes users from the loop should be redesigned as collaborative interaction using Canva's subjective-language input pattern as reference

  2. Create a data instrumentation ticket to capture user edit histories and revision sequences in a model-training-ready format by end of Q2

  3. Establish a 'reliability score' KPI for every AI feature — track error rate, hallucination rate, and user-reported inaccuracy — and set a threshold below which you won't ship new capabilities

  4. Conduct a 'last-mile gap' analysis for your product category by end of Q2: where do ChatGPT/Claude/Gemini generate output that users struggle to make professional and publish-ready?

GPU Costs +50%, Manus at $2B, Fine-Tuning Barriers Gone: Your AI Architecture Calculus Just Flipped

The GPU Tax Is Real and It's Breaking Roadmaps Now

GPU prices are up ~50%, driven by AI agent demand so intense it's causing service outages at major providers and forcing actual product cancellations. This isn't a supply chain hiccup — it's a structural shift in the cost basis for AI products. If you spec'd AI features with 2025 infrastructure pricing, those assumptions are broken. The immediate action isn't strategy — it's pulling up your unit economics spreadsheet and running the numbers at 1.5x GPU cost. Some features will survive; others won't.

Competitor product cancellations from the GPU cost spike create acqui-hire and talent acquisition opportunities — if you're positioned to move fast.

Meta Proved the Harness Is Worth $2B — Not the Model

Meta's acquisition of Manus for ~$2B specifically targeted its agent harness technology: memory management, skill encoding, protocol handling, observability, evaluation loops, and compression. Not the underlying model. Boris Cherny (creator of Claude Code) claims evaluation alone drives 2-3x quality improvements. The harness taxonomy — Memory (working context, semantic knowledge, episodic experience), Skills (procedures, heuristics, constraints), Protocols (agent-to-user, agent-to-agent, agent-to-tools) — is the clearest architectural framework for scoping agent features available.

If you've been allocating 80% of your AI investment to model selection and prompt engineering, and 20% to the infrastructure around it, you likely have the ratio backwards.

The Fine-Tuning Barrier Just Collapsed

GRPO (the algorithm behind DeepSeek-R1) combined with RULER (LLM-as-judge for reward signals) eliminates the two biggest barriers to custom model training: you no longer need hand-crafted reward functions or labeled datasets. The ART framework packages this with LangGraph/CrewAI integrations and vLLM inference. The practical implication: if you can articulate 'what good looks like' for your agent's task — even qualitatively — you can train a specialized model that outperforms GPT/Claude on that specific task at a fraction of the cost.

The Hybrid Architecture Is No Longer Optional

Simultaneously, the on-device pipeline has reached production-readiness: Qwen3-0.6B runs at ~25 tokens/s on iPhone 17 Pro via a 470MB artifact using ExecuTorch — already in production across Instagram, WhatsApp, and Messenger. The cost-quality frontier demands a three-tier approach:

  1. On-device (ExecuTorch, 470MB): privacy-sensitive, latency-critical, zero-inference-cost features
  2. Fine-tuned small models (GRPO+RULER, 3B params): high-volume specific tasks currently on expensive cloud APIs
  3. Frontier cloud (Opus 4.7, GPT-5.4): complex reasoning where no alternative matches quality

The 50% GPU price increase makes this three-tier shift from 'nice to have' to 'required for viable unit economics.'

What to do

  1. Re-forecast all AI feature unit economics at 1.5x GPU cost this week — identify which features remain viable and which need architectural rearchitecture or cancellation

  2. Map your agent architecture against the harness taxonomy (Memory/Skills/Protocols) and identify your two biggest investment gaps by end of sprint

  3. Identify your top 3 API-dependent features where a fine-tuned 3B model could replace cloud inference, and assign a spike to benchmark GRPO+RULER on your highest-volume task

  4. Prototype LLM-as-judge evaluation using relative ranking ('which is best among 4?') for your highest-volume AI feature

CISOs Are Blocking Your AI Rollout — Here's What Makes the Security Questionnaire

Shadow AI Is the #1 CISO Pain Point — And Every One Sounds 'Defeated'

Q1 2026 CISO conversations reveal a consistent pattern: shadow AI is the dominant security concern, and every security leader interviewed sounded 'a little defeated about it.' Enterprise AI rollouts (Copilot, Gemini, ChatGPT Enterprise) are being blocked or rolled back because product teams ship without ACL cleanup, agent inventory, or prompt injection testing. If your product sells to enterprises with AI features, this is now a top-4 line item on security evaluations.

The ACL Debacle: A Cautionary Tale for Every AI Feature Touching Enterprise Data

The most concrete failure: Microsoft Copilot faithfully surfaces board decks to interns because SharePoint has a decade of stale access control lists. CISOs describe deploying without ACL cleanup as a 'self-inflicted breach.' The same pattern applies to any AI feature querying existing enterprise data stores. The winning move is to make permissions auditing part of your onboarding flow — turn the buyer's biggest objection into your product's differentiator.

Blanket blocking of AI tools drives usage underground — onto phones, personal devices, and mobile hotspots — making the security problem worse, not better. CISOs know this, which creates massive demand for products that enable controlled adoption.

Three Attack Vectors Converging

VectorStatusYour Exposure
Prompt injectionCalled 'the new SSRF' by CISOsAny AI feature reading emails, tickets, PDFs, web pages
AI supply chainCoding tools hallucinate package names; attackers squat themEvery engineering team using Copilot/Cursor/Claude Code
LLM guardrail bypassWharton: persuasion techniques 2x+ bypass rateAny AI feature with safety-critical outputs

Johns Hopkins' MANYIH-BENCH research adds another dimension: current frontier models cannot reliably resolve conflicts across multiple privilege levels. This matters for any product where an AI agent operates across different access tiers. Don't rely on prompt-level instructions for access control — enforce it at the application layer with hard gates.

The Opportunity Side

The identity security vendor landscape is fragmented — each vendor tackles one slice (infostealer monitoring, MFA, non-human identity, session security) with no platform player. CISOs are overspending on unintegrated tools and doing custom engineering. The 2026 CISO priority stack explicitly allocates AI security as its own line item with dedicated headcount. Products that ship governance controls (sanctioned AI tiers, usage dashboards, granular role-based access) close enterprise deals that all-or-nothing competitors lose.

What to do

  1. Audit your own team's use of autonomous AI agents (Claude Code, Cursor, Zapier AI) running against production systems with personal tokens this week — create an inventory and advocate for scoped service accounts before security imposes a blanket ban

  2. Add prompt injection testing to your AI feature QA process as a P0 requirement alongside XSS and SSRF testing, especially for features ingesting external content

  3. If shipping any AI feature that queries enterprise data, add a permissions/ACL health-check to your onboarding flow and document it in your PRD as a hard prerequisite

  4. Verify your engineering team pins GitHub Actions to commit SHAs (not tags) and has install-time package verification for all dependencies, especially AI-suggested ones

The bottom line

GPU costs are up 50% and breaking AI roadmaps, Meta just priced the agent orchestration layer at $2B (not the model), and the two largest AI user studies ever conducted — Canva's 265M users and Anthropic's 81K-person survey — both say the same thing: reliability and collaboration beat raw capability. Meanwhile, CISOs are actively blocking enterprise AI rollouts over ACL failures, prompt injection, and agent sprawl. The product teams that win Q3 aren't the ones shipping the most AI features — they're the ones shipping reliable, collaborative AI with hybrid cost architectures and security that survives the enterprise questionnaire.