Get tomorrow's brief in your inbox
Date: April 30, 2026
Today: Big Tech's Q1 earnings reveal a compute crunch limiting AI growth, with Google Cloud up 63% but unable to meet demand. n8n shipped an MCP server that lets you build workflows from prompts directly in Claude or ChatGPT. Anthropic is reportedly fielding $50B fundraising offers at a $900B valuation as revenue hits a $40B run rate.
The AI Compute Bottleneck Is Here
Every major cloud provider just told Wall Street the same story: they cannot build data centers fast enough to meet AI demand. Google Cloud grew 63% year-over-year to hit $20 billion in revenue but CEO Sundar Pichai admitted "we are compute constrained in the near term" and that "our cloud revenue would have been higher if we were able to meet the demand." Google's cloud backlog doubled to $462 billion. Google Cloud surpasses $20B
Amazon Web Services grew 28%, its fastest clip in 15 quarters, hitting $37.6 billion. Microsoft is now forecasting $190 billion in capital spending for 2026, up from earlier estimates, blaming a $25 billion hit from higher memory component prices. Meta is projecting $125-145 billion in capex, also citing memory pricing pressure. Microsoft calls for $190 billion in 2026 capital spending | Amazon's cloud business is surging
If you're planning to launch AI-dependent products or scale existing ones, factor in potential delays accessing compute. OpenAI's recent multi-cloud expansion (adding AWS alongside Azure) is a direct response to this constraint. Small businesses should lock in cloud commitments now if your roadmap depends on model access, and consider building on providers with spare capacity or multi-cloud fallback plans.
n8n's MCP Server Lets You Build Workflows from Prompts
n8n released an MCP server that connects to Claude, ChatGPT, Cursor, or any MCP-compatible AI client and lets you describe a workflow in plain language. The AI builds it directly in your n8n instance, validates it, runs test executions, and fixes errors on its own. You can iterate conversationally: "make it less code-heavy" or "use the Gmail node instead of a code block." Works in all editions: Cloud, Enterprise, and the free self-hosted Community Edition. Build and Update Workflows with n8n's MCP Server
If you use n8n for automation, this cuts workflow development time dramatically. For teams without dedicated automation engineers, it lowers the skill floor. You still need to know what you want automated, but you no longer need to wire the nodes yourself.
Simon Willison's LLM 0.32a0 Refactor
Simon Willison released an alpha of his LLM Python library with a major change: prompts are now modeled as a sequence of messages (matching the OpenAI chat completions API format) instead of single text inputs. This makes it easier to feed in prior conversations, build chatbots, or emulate vendor APIs. LLM 0.32a0 is a major backwards-compatible refactor
If you're building custom LLM tooling in Python, this library now handles multi-turn conversations cleanly without needing to roll your own SQLite storage layer.
Human-in-the-Loop vs. Human-on-the-Loop: When to Use Each
n8n published a breakdown of when to require human approval before AI acts (human-in-the-loop / HITL) versus letting AI run autonomously with humans reviewing results (human-on-the-loop / HOTL). HITL is synchronous: AI stops and waits for approval. Use it for high-stakes actions like customer emails, financial transactions, or compliance-regulated decisions. HOTL is asynchronous: AI executes, logs anomalies, and humans review later. Use it for high-volume, lower-risk tasks like product description generation or routine order processing. Human-in-the-Loop vs. Human-on-the-Loop
The article recommends hybrid setups: let AI generate 1,000 product descriptions autonomously (HOTL) but require approval before publishing the top 50 to your homepage (HITL). This scales human attention to where it matters most. If you're deploying AI agents, map your workflows to this spectrum before going live.
The Zig Project's Anti-LLM Contribution Policy
The Zig programming language project bans all LLM-assisted code contributions, including GitHub issues, pull requests, and bug tracker comments. VP of Community Loris Cro explains: Zig values contributors over contributions. The goal of reviewing PRs is not to land code, it's to grow trusted, long-term contributors. LLM assistance breaks that. A perfect PR written by an LLM teaches the Zig team nothing about the human behind it. They call it "contributor poker": you bet on the contributor, not the contents of their first PR. The Zig project's rationale for their firm anti-AI contribution policy
If you maintain open source projects or manage engineering teams, this is worth considering. The tradeoff: slower short-term code velocity for higher-quality long-term contributors. It also raises a question: if a PR was mostly written by an LLM, why should a maintainer spend time reviewing it instead of firing up their own LLM to solve the same problem?
Microsoft Copilot Crosses 20 Million Paid Seats
Microsoft said it now has over 20 million paid Copilot users (up from 15 million in January) and that weekly engagement is now at the same level as Outlook. The $30/user/month add-on for Microsoft 365 is becoming a habit for enterprise users. Microsoft says it has over 20M paid Copilot users
If you're a Microsoft 365 shop and haven't tried Copilot yet, the usage data suggests it's delivering value. Consider a pilot with your most document-heavy teams (sales, legal, customer success) before rolling out company-wide.
OpenAI Expands to AWS, Ends Microsoft Exclusivity
OpenAI made its models available on Amazon Bedrock (AWS's AI service) and expanded its AWS spending commitment by $100 billion. Microsoft no longer holds exclusive cloud provider status for OpenAI's API products. This follows a broader restructuring where Microsoft lost its exclusive license to OpenAI's IP and its revenue share agreement ended. OpenAI's subtle drift from Microsoft has become an aggressive move toward Amazon | Amazon's cloud unit reports 28% sales growth
If you're locked into AWS and wanted to use OpenAI models without routing through Azure, you can now do that natively. Expect pricing and latency improvements as AWS and Azure compete for your workload.
Anthropic Fielding $50B Fundraise Offers at $900B Valuation
Sources say Anthropic is considering a $40-50 billion raise at a $850-900 billion valuation, which would more than double its February valuation of $380 billion. The company's revenue run rate has hit $40 billion (up from $9 billion at the end of 2025), driven heavily by Claude Code and enterprise coding tools. A board meeting in May will decide whether to proceed. Sources: Anthropic could raise a new $50B round
For small businesses, this signals continued investment in developer-facing AI tools. Anthropic's Claude Code and API products are likely to see more features, better pricing, and expanded capacity. If you're building on Claude's API, expect the platform to keep improving.
Nvidia Invests in Legal AI Startup Legora
Nvidia's venture arm backed Swedish legal AI company Legora at a $5.6 billion valuation as part of a $50 million Series D extension (total round: $600 million). Legora builds AI agents to automate legal workflows and just crossed $100 million in annual recurring revenue serving law firms like White & Case and corporate legal departments at Barclays. It's Nvidia's first bet in legal tech. Nvidia just invested in the AI legal startup
If you run a service business with heavy contract review, legal research, or compliance work, legal AI tools are maturing fast. Legora's revenue growth suggests enterprises are actually using these tools, not just piloting them.
OpenAI's "Goblin" Problem and How They Fixed It
OpenAI published a case study on how GPT-5.1 and later models started using words like "goblin" and "gremlin" in metaphors at a 175% higher rate. The root cause: training for the "Nerdy" personality (used by 2.5% of users) inadvertently rewarded creature metaphors, and the behavior transferred to all outputs. They tracked the reward signal, found it boosted goblin-heavy responses 76.2% of the time, and the style leaked across the model. Where the goblins came from
This is a fascinating look inside model training. If you're fine-tuning models or building custom personalities, it's a reminder that subtle reward signals can propagate unexpectedly. Test broadly, not just on the target use case.
CRM Decision Framework
If you're choosing between HubSpot and Pipedrive: Pipedrive is better value for pure sales teams ($14/user/month for the base plan), HubSpot is worth it if you need marketing, support, and content tools to connect natively across departments. Both now have AI prospecting agents. HubSpot's free CRM is still the best starting point for small teams. Pipedrive vs. HubSpot: Which CRM is right for you?
Big Tech's compute crunch is real and will define the next 12-18 months of AI product launches. If your roadmap depends on cloud AI, assume longer lead times and higher costs. The simultaneous collapse of Microsoft's OpenAI exclusivity and OpenAI's expansion to AWS signals a multi-cloud future for frontier models. Expect more vendors to offer the same models across providers, making infrastructure choice more important than model choice. Watch for smaller, more efficient models to gain share as businesses look to escape the compute bottleneck.