The Product Desk

The tell is the pivot, not the price. A platform team probably wrote "open source stays free as the fallback" into last quarter's cost model. Meta built its AI credibility on open-source Llama and just walked away to sell proprietary API access, so that line is now wrong. Zuckerberg faces explicit Wall Street pressure to monetize AI infrastructure without a cloud business. Developers pay. The free option narrows and the paid floor drops in the same move.

The economics here are the part worth acting on. Muse Spark 1.1 scores 51 on the Intelligence Index, tied with GPT-5.6 Luna, with a 1M-token context window and 114 tok/s median speed. For summarization, extraction, classification, and most non-frontier production tasks, this is the new reference price:

The window is time-boxed, and that changes the negotiation. OpenAI and Anthropic are preparing IPOs. Pre-IPO vendors chasing logo retention will cut deals now that public-market margin discipline won't allow later. Google is running the parallel play, giving Gemini Live's multimodal planning away free globally to anchor expectations at zero. Between Meta subsidizing from a $2B/day ad business and Google subsidizing for platform pull, sustained deflation is the base case for two quarters. Sequoia warns this spending needs revenue that doesn't yet exist, which is the reason to lock in flexibility, not multi-year commitments.

One caveat separates the pitch from the thing being done. Meta's benchmark claims, beating Opus 4.8 and GPT-5.5 on tool use and financial analysis, are self-reported. Validate on your own workloads before migrating anything customer-facing.

Meta's cut removes open source as a free fallback and replaces it with a paid floor cheap enough to reprice your whole backlog. That is the decision on the table, not the headline price.

On July 1, Sysdig documented JADEPUFFER. A human picked the target. After that an LLM ran reconnaissance, exploitation, lateral movement, and encryption of a production database on its own, via Langflow CVE-2025-3248 and Nacos CVE-2021-29441, with payloads carrying natural-language reasoning annotations no human operator writes. What changed is speed. Threat models that assume human-paced lateral movement and hours of dwell time are now wrong. If the safety story depends on detection-before-damage, the clock moved from hours to minutes. A second actively exploited Langflow CVE, CVE-2026-55255 at CVSS 8.4, is harvesting credentials now. For teams that adopted Langflow, that is a today problem.

The defense side reads worse than the offense. Here is what teams tell themselves: agents run under managed, scoped identities. Here is what teams actually do: run agents on shared credentials (69%), assign unique identities only 32% of the time, and 54% have already eaten an agent incident. The Amazon Bedrock gateway attack shows why that compounds. AI gateways centralize identities, permissions, and model access into one blast radius, and attackers took one with a familiar cloud playbook, no novel techniques. We have watched the same shared-credential mistake with service accounts for years. A multi-vendor vulnerability across Amazon, Anthropic, Google, and Cursor tooling let agents feed false information to human approvers, which breaks the 'human reviews everything' assumption most PRDs lean on.

The product read is simpler than it looks. JetBrains launched a governance suite above Claude, Codex, Gemini, and Junie: access controls, usage visibility, cost management. A 6-stage non-human-identity maturity model is circulating because standard IAM can't govern agents. Put those together and the governance layer is being built now. Security questionnaires will ask about per-agent identity and prompt-injection boundaries within a quarter or two. Here is the forcing function. On one axis: does an agent carry its own identity, or share one. On the other: is its permission scope logged and verifiable, or assumed. Ship per-agent identities, scoped permissions, and verifiable audit trails, and deals close that competitors can't touch. Ship ungoverned agents and incident liability accrues instead.

The first AI-executed ransomware operation ran end-to-end while most enterprises still can't tell their agents apart. Agent identity governance is a deal-closing feature now, not a security afterthought.

The dangerous part isn't the numbers — it's that they're invisible in your dashboards. 'Ghost citations' mean AI systems recommend products without trackable referral events, so attribution-driven roadmaps systematically undervalue the highest-converting discovery channel that exists. Weight channels by measured conversion and you misallocate spend toward channels you can see, away from the one actually working.

The structural problem compounds it: AI systems cite premium third-party content 3x more than brand-owned pages. A decade of SEO playbook — build owned content, rank it, harvest traffic — inverts. Earned media in AI-trusted publications beats your own blog, making content a product strategy problem: your citation footprint determines whether AI assistants recommend you.

The market has voted with headcount. AEO (AI Engine Optimization) is splitting from SEO into a standalone function — 50+ senior roles open at leadership-level salaries against a tiny talent pool. Companies are building permanent teams, not experimenting. It's SEO's 2015 moment, except the discipline is forking rather than maturing — the talent arbitrage window is short.

A working measurement pipeline is implementable this sprint: Google Search Console's new Generative AI report shows which pages appear in AI Overviews; pair it with Screaming Frog for content structure and Claude Sonnet to analyze which prompt patterns trigger your citations. It won't capture ghost citations — nothing does yet — but it turns a dark channel into a partially instrumented one.

The highest-converting acquisition channel in your funnel is one your analytics can't see and your own content can't win — earned citations just became a product surface.

Watch what a team actually does with Gamma before you argue about the valuation. Fifty million people use it. Six hundred thousand pay. That is 50M users, 600K paying, a 1.2% conversion that only prints because CAC is zero. The output is the distribution. Every shared deck recruits the next user, so the product does the acquisition the marketing team keeps putting in the deck. Lovable's $3.4M revenue per employee is the same story from the build side: tiny teams shipping what used to take departments. Here is what a board should ask, not what it usually tells itself: if a competitor can replicate your feature set with 10% of your headcount, what's your moat? "We have more engineers" used to be a line in the pitch. It is now a line item on the liability side.

The pricing shift is the part most decks skip. Separate the thing being sold from the thing being done. An agency selling outcomes instead of tools captures 95-100% of a client's OPEX for a problem. A software vendor captures 5-10%. That is a 10-20x revenue multiple per account, and the agency runs it at software margins because the work is automated. If a product is priced per-seat in a domain where AI can deliver the outcome — support, content, QA, analytics — someone is already modeling how to bill the customer's whole team budget instead of the license fee.

Now hold that against the demand side, because today's sources disagree and both are right. Sequoia's David Cahn puts a number on the gap: AI needs $3 trillion in revenue to justify 2026 spending. Meanwhile Starbucks is ripping out Oracle and IBM to build in-house, which is a buyer concluding it can make what the vendor sells. AI-native supply is hyper-efficient and demand is skeptical and self-sufficient. Value collects around zero-CAC distribution loops and outcome-priced delivery. Undifferentiated per-seat tools get squeezed from both ends: efficient startups below, insourcing buyers above.

The market is splitting into products that grow themselves and products that get built in-house — per-seat tools without a distribution loop are caught in the middle.