Every week I read the major AI sources and write down what actually matters for people building production systems. This week’s signal was hard to miss: the supply side accelerated at a pace that makes the enterprise side look like it’s standing still.
$65B in a single funding round. Coding agents from three labs shipping in the same week. An 8-month legacy migration compressed to 8 days. And on the other side, Gartner warning that 40% of deployed agents will be decommissioned within two years because nobody built the governance layer.
I spent the week building exactly that layer for our own AI practice. R&D charter, case study system, decision framework for what gets funded and what gets killed. Here’s what happened this week.
THE BIG FOUR
1. Anthropic Raises $65B at $965B
Anthropic closed a $65B Series H at $965B post-money, surpassing OpenAI’s $730B from February. Revenue hit $47B annualized, up from $14B in February. That’s 3.4x in three months.
The round includes $15B in previously committed hyperscaler investments. Claude is now natively on AWS Bedrock, Google Cloud Vertex AI, and Azure: the first frontier model on all three major clouds. Separately, Anthropic signed a $1.25B/month compute deal with SpaceX’s Colossus infrastructure (through May 2029, 90-day exit clause).
Why it matters: $47B annualized means Claude is a production platform, not a research project. Multi-cloud availability removes the last lock-in objection for enterprise procurement. The SpaceX compute deal signals inference demand outstripping hyperscaler capacity.
2. Coding Agents Go Three-Front: Opus 4.8, Devin $1B, Grok Build
Three major coding agent releases in one week. Anthropic launched Claude Opus 4.8 with Dynamic Workflows enabling up to 1,000 parallel subagents per session (16 running concurrently). SWE-Bench Verified hit 88.6%. Fast Mode at $10/$50 per million tokens, 3x cheaper than standard. Just 41 days between Opus 4.7 and 4.8.
Cognition raised $1B at $26B valuation for Devin. Their engineering team now commits 89% of code through Devin, up from 13% in December. Revenue grew from $37M to $492M run-rate in 12 months. Mercedes-Benz used Devin to cut an 8-month COBOL modernization to 8 days (per Cognition’s blog and a joint Mercedes-Benz announcement, covering 200,000+ lines of code). Worth noting: both sources are Cognition marketing materials. xAI launched Grok Build, a terminal coding agent with plan-before-execute mode and a Skills extension system.
Why it matters: If the Mercedes-Benz figure holds up under scrutiny, 8 months to 8 days on legacy COBOL is not incremental productivity. It is a category collapse in project timelines. Devin writing 89% of Cognition’s own code is the credibility proof: the company building the agent trusts it with its own codebase. Three labs shipping in one week makes coding agents the most competitive category in AI right now.
3. Enterprise Agent Decommission Warning: 40% by 2027
The supply shock has a mirror image. Last edition covered the Sinch data (74% of enterprises rolling back AI agents, 2,527 decision-makers, 10 countries). This week added more weight.
Gartner formally warned that binary governance (locked-down or fully-trusted) is the root cause, predicting 40% of autonomous agents decommissioned by 2027. Their recommendation: a 4-level proportional autonomy model (Observe, Advise, Act with Approval, Act Autonomously). Check Point found a 51-point gap between AI security strategy (77% updated) and enforcement (26% can actually enforce), with 78% reporting confirmed or suspected AI security incidents.
Why it matters: 40% decommissioned is not a minor correction. It means nearly half the autonomous agents enterprises have deployed will be pulled back because the governance layer was never built. The Check Point data explains why: 77% updated their security strategy on paper but only 26% can enforce it. Strategy without enforcement is a press release, not a control.
4. Governance Stack and Regulation: Two Responses
OpenAI shipped Skills permissions and a Compliance Logs Platform to ChatGPT Enterprise (7M seats). Target: the $670K shadow-AI breach cost premium IBM identified. 92% of Fortune 500 companies use ChatGPT. Most had no way to control what employees did with it.
The EU reached provisional agreement on the Digital Omnibus on AI. Article 50 transparency obligations hold at August 2, 2026 (66 days). The U.S. pulled its AI safety executive order on May 22. More than 2,000 AI-related state bills are active across US legislatures.
Why it matters: OpenAI shipping governance tooling to 7M seats signals the end of the “just trust the users” era. The regulatory trajectory is EU-first, US fragmented. European enterprises have 66 days until the first binding transparency requirement.
ALSO WORTH KNOWING
-
• AI coding productivity: gains concentrate at the top. Cursor’s Developer Habits Report shows lines of code per developer per week went from 3.6K to 8.6K in 18 months. But top 1% of developers now produce 46x more code than the median. One unnamed 1,000-engineer company on Claude Code reports organizational gains of “1+1+1+1=1.5”. A 400-engineer SaaS company cut PR-to-production from 4.2 days to 6.4 hours using a 5-agent Claude Code pipeline.
-
• Snowflake signs its largest deal ever: $6B multi-year AWS commitment framed around the “agentic enterprise.” Up from $2.5B in 2023 and $1.2B at IPO.
-
• Anthropic launches a PE-backed enterprise services firm with Blackstone, Goldman Sachs, General Atlantic, Apollo, GIC, and Sequoia. Target: mid-market companies that need Claude deployed but lack the engineering capacity.
- • Anthropic previews Mythos-class models and Glasswing security framework. Glasswing has identified more than 10,000 vulnerabilities across open-source codebases. Both signal Anthropic expanding beyond the core model layer.
- • MCP protocol architecture overhaul. Next release candidate moves toward stateless tool calls, deprecates Roots, Sampling, and Logging in favor of OpenTelemetry. Final spec expected July 28.
- • Mistral rebrands Le Chat as Vibe (free/Pro at $14.99/Team at $24.99), a unified work-and-code agent platform with a 10 MW dedicated inference facility in Les Ulis, France.
THE PATTERN
• Funding: Anthropic $65B Series H at $965B valuation.
• Compute: G$1.25B/month SpaceX Colossus deal (Anthropic).
• Coding agents: Opus 4.8 (1,000 subagents), Devin $1B raise (89% self-coded), Grok Build (xAI).
• Governance response: 90% Gartner 40% decommission prediction, 51-point enforcement gap , OpenAI governance tooling to 7M seats.
• Regulation: Anthropic’s $10.9B Q2 revenue and $559M profit. Two years early. The market is consolidating around the winners.
• Infrastructure: Snowflake $6B AWS deal, Anthropic PE-backed services firm, Mistral 10 MW inference facility.
• Protocol: MCP stateless overhaul, final spec July 28.
The supply side shipped $65B rounds, $1.25B/month compute deals, and coding agents that compress 8-month projects to 8 days. The enterprise side is still building the governance layer for what it already deployed. I spent the week on the same problem at a smaller scale. The R&D function I designed for our AI practice has one job: decide what gets built and what gets killed, fast enough to stay relevant, slow enough to avoid waste. The charter was straightforward. The kill criteria were not. What does “not working” look like three weeks into a prototype? Most teams never define that. So nothing ever dies. And the backlog becomes the strategy.
That is the real story this week. The technology moves at $47B run rates and 41-day release cycles. The organizational capacity to govern it moves at quarterly review speed. The gap is widening.
Sources: Anthropic press release, SpaceX S-1 filing, Sinch survey, Gartner, Writer, Deloitte, Check Point, Cognition blog, Mercedes-Benz/Cognition announcement, Cursor Developer Habits Report, OpenAI blog, IBM Cost of a Data Breach Report, Gibson Dunn, European Commission, CIO.com, TechCrunch, Business Insider, Axios, AlphaSignal, The Rundown AI, Prohuman AI, Exponential View, The Batch.
I write about Production AI, enterprise AI adoption, and building systems that actually work. Follow along if that’s your thing.
Latest Articles
Discover materials from our experts, covering extensive topics including next-gen technologies, data analytics, automation processes, and more.
