The Product Desk

$285B Evaporated. Here's What Actually Got Repriced.

Anthropic's Cowork launch didn't just move markets — it repriced entire software categories. Investors wiped $285B in SaaS market cap, coining 'SaaSpocalypse,' and shifted capital toward products that own data and workflows over surface-level AI wrappers. The trigger: Cowork ships with 11 open-source plugins that natively connect to Salesforce, Snowflake, BigQuery, Jira, Linear, Notion, Zendesk, Intercom, Slack, and HubSpot — your entire operational stack — installable in 30 seconds. Six third-party libraries (Skills.sh, SkillsMP, Smithery, SkillHub, and two official directories) already host thousands of pre-built Skills. Partner integrations from Asana, Atlassian, Canva, Figma, Sentry, and Zapier are live.

If your product is essentially 'connect Tool A to Tool B and add a dashboard,' Claude just replicated your value proposition with an open-source plugin. The market just told you it agrees.

The Skills architecture uses an open standard — SKILL.md files with YAML frontmatter, portable across Claude Web, Claude Code, and Cowork. Anthropic published a 32-page technical guide and shipped a meta-skill that creates new skills automatically. This is ecosystem acceleration by design: the AI equivalent of Apple launching the App Store with a developer toolkit, except everything is open-source and the ecosystem self-replicates.

Atlassian's Counter-Evidence: Opacity Kills, Transparency Wins

Here's where the narrative gets nuanced. In the same week, Atlassian's CTO revealed they scrapped and rebuilt their Rovo Dev AI agent after internal engineering teams refused to use the original 'one-click magic' version. The output was useful. Developers rejected it anyway — because they couldn't see the agent's reasoning. This is the most concrete first-party case study on the transparency-vs-automation tradeoff in production.

Post-rebuild metrics tell the story: 45% reduction in PR cycle time, 51% of security vulnerabilities auto-resolved, and developer satisfaction jumped from 49% to 83%. The key design principle: every AI-assisted decision must have a clear human owner. If AI behavior cannot be understood or observed, it doesn't belong in a critical path.

Atlassian's CTO also delivered the sharpest counter to the SaaSpocalypse thesis: customers buy workflows, compliance, shared context, and reliability — not just code. An AI can mimic your UI in a weekend but can't replicate your compliance certifications, your data model encoding years of domain knowledge, or your team's shared understanding of how work flows. This tracks with a critical observation from multiple sources: most AI products are single-player (Copilot helps one dev, ChatGPT helps one writer). The competitive whitespace is multiplayer AI — shared agent sessions, cross-functional context synthesis, collaborative planning.

Where This Leaves Your Product

The tension between these two data points — $285B destruction and Atlassian's 83% satisfaction — isn't a contradiction. It's a sorting mechanism. Products in the kill zone: workflow orchestration layers without proprietary data, 'connect and dashboard' tools, surface-level AI wrappers. Products that survive: those with deep team context, compliance moats, inspectable AI reasoning, and data network effects. The 45% of AI-generated code that still contains security flaws makes Atlassian's guardrails-first approach a competitive advantage, not an overhead cost.

Concurrency Collapse: The Hidden Tax on Every Long-Context Feature

The most consequential infrastructure number this week isn't a model benchmark — it's a cost curve. A 70B parameter model on an H100 serves 59 concurrent users at 4K context, generating 7.4M output tokens/hour at $0.34/M tokens. Stretch that to 128K context: 1 user, 126K tokens/hour, $19.84/M tokens. That's a 58x cost increase for a 32x context increase. The KV cache for a single 128K session consumes 20.97 GB (INT8 quantized) — the GPU simply runs out of room for anyone else.

At 128K context, raw hardware cost exceeds what OpenAI and Anthropic charge at retail. If you're self-hosting long-context inference without architectural optimization, you're paying more than the frontier labs charge.

Every PRD that says 'supports 128K context for document analysis' has an implicit infrastructure cost most teams don't model correctly. The O(n²) attention computation during prefill makes the math worse, not better, as you scale.

Three Tiers of Solutions — With Real Tradeoffs

MLA (Multi-head Latent Attention) is the highest-leverage near-term optimization: DeepSeek-V2's approach achieves 93.3% KV cache reduction, restoring concurrency to 27 users at 128K. Hybrid architectures like Jamba (1 attention layer per 7 Mamba layers) reduce cache by 87% and fit a 50B MoE at 256K on a single H100 — but require 2-4 month serving stack rewrites because vLLM's PagedAttention assumes KV-cache-only state. Kernel switching overhead eats 10-15% of theoretical savings.

The tactical move: segment your long-context features by recall pattern. Exact retrieval needed (legal, code) → full attention/hybrid. Recency-dominant (chat, streaming) → StreamingLLM (22.2x speedup, zero memory growth, but permanently loses anything outside rolling window). Summary-sufficient (document overview) → pre-summarization pipeline. Each maps to a different cost profile.

The Edge Gap Is a Separate Problem

Current data center architectures completely fail on mobile devices with 4-6 GB RAM and 1/500th H100 bandwidth. Liquid AI (continuous-time networks), xLSTM (gated recurrences), and RWKV (mobile-scale linear attention) are building edge-specific architectures. Meanwhile, Inception's Mercury 2 diffusion-based LLM hints at a paradigm that could bypass transformer limitations entirely. For roadmaps beyond 12 months, keep your serving stack modular enough to swap architectures without rebuilding everything.

The Qwen Implosion Is a Vendor Risk Event

Within 24 hours of shipping Qwen 3.5 — the open-weight model that community reports put at Sonnet-level performance — technical lead Junyang Lin and two key researchers abruptly resigned. This is the third senior departure from Qwen's team in 2026. The root cause: Alibaba is reorganizing the vertically integrated research team into horizontal, KPI-driven units focused on DAU growth. It's the classic pattern of corporate short-termism destroying research excellence, and it triggered a 5.3% drop in Alibaba's Hong Kong shares.

With 600M+ downloads, Qwen is embedded in countless production systems. If model quality degrades or release cadence slows, those systems need alternatives — and the alternatives aren't ready.

The irony: Qwen 3.5 is genuinely impressive. The 9B model outperforms OpenAI's 120B open-source model on graduate-level reasoning while running on 8GB GPU with 4-bit quantization. The 397B sparse MoE activates only 17B parameters per token, delivering Sonnet-level quality at a fraction of inference cost. But the team that built it is disintegrating. If you've built production features on Qwen, create a documented fallback plan this quarter.

The Western Open Frontier Gap Is Confirmed

Reflection AI's CTO — a founding DeepMind engineer who led Gemini post-training — pivoted his entire company from shipping a coding agent to building the 'missing Western open frontier model.' His reasoning: 'Llama 4 is not a particularly strong model' and 'the whole Western ecosystem was missing a powerful open base model.' This conviction attracted a $20B valuation with zero shipped products, zero published research, and a fully pivoted strategy — a ~37x valuation increase in one year.

Whether Reflection delivers is uncertain. What's confirmed is the gap: Chinese labs (DeepSeek, Qwen) dominate open frontier models, creating a provenance problem for any product serving government or regulated enterprise customers. Sovereign AI demand is emerging as a distinct market segment willing to pay premium for Western-origin models. If you haven't asked your enterprise customers about model origin preferences, add it to your discovery script.

Self-Hosting Economics Shifted — But So Did the Risk

The GPU monopoly is cracking: Meta validated AMD MI300 for production LLM inference, open-sourcing RCCLX with meaningful gains in decode latency. Combined with Qwen 3.5's efficiency (17B active params for frontier quality) and tools like Unsloth enabling fine-tuning of the full Qwen family at 1.5x speed and 50% less VRAM, self-hosting is economically viable for the first time at frontier quality. But the Qwen team implosion introduces a new risk dimension: open-model dependency now carries organizational risk, not just technical risk. You need a model-switching runbook, not just a model evaluation matrix.