Sequoia Capital partner Julien Bek published "Services: The New Software," arguing that for every dollar spent on software, six go to services, and the next trillion-dollar AI company will capture that services budget by selling outcomes directly. The fastest-growing AI companies in 2025 were Copilots. In 2026, the leaders will be Autopilots: billing for the task completed, not the user enabled. A companion piece, "From Hierarchy to Intelligence," argues AI now eliminates middle management as an information-routing function.

⮕ The seat-license pricing model is losing its rationale as AI agents take over the underlying tasks. If you are evaluating software vendors today, the most diagnostic question is no longer "how much per seat?" but "will you do this task and bill me per output?" That question tells you which vendors have actually adapted.

NVIDIA announced a $2 billion investment in Marvell Technology to develop NVLink Fusion, an interconnect architecture that lets non-NVIDIA accelerators communicate with NVIDIA's GPUs, CPUs, and networking hardware. Marvell shares rose 7% on the announcement. The move opens NVIDIA's previously closed NVLink ecosystem to third-party chips, while requiring that every NVLink Fusion platform include at least one NVIDIA component.

⮕ This is an infrastructure layer play, not a chip investment. NVIDIA is building the standard through which all accelerators, including those from competitors, must connect. For teams building AI infrastructure, ecosystem lock-in is now being designed at the interconnect level, not the model or software layer.

PrismML, a Caltech-backed startup, launched Bonsai on March 31: a 1-bit model that fits 8 billion parameters into 1.15GB, 14 times smaller than full-precision equivalents, while matching competitive benchmarks. On iPhone it runs at 40 tokens per second. On an RTX 4090, 440. Caltech holds the underlying IP; PrismML is the exclusive commercial licensee.

⮕ If 1-bit quantization consistently delivers near-full-precision performance at 14x compression, AI inference economics shift from cloud-dependent to edge-native. For teams in healthcare, legal, or finance where data cannot leave the device, this is a capability unlock that has not existed at this quality level before.

Salesforce launched 30 new AI features for Slack, including autonomous agents that execute multi-step tasks without leaving the application. Slackbot now functions as an MCP client, connecting to outside services including Salesforce's Agentforce platform. This is the company's clearest bid yet to turn Slack from a communication tool into the surface where AI agents and humans share the same workspace.

Investors deployed $297 billion globally in Q1 2026, more than any previous quarter on record, according to Crunchbase. Four frontier labs captured 64% of that total: OpenAI ($120B), Anthropic ($30B), xAI ($20B), and Waymo ($16B). The investment cycle is outpacing productization at a pace that has no historical parallel.

Mitchell Katz, president and CEO of NYC Health + Hospitals, said at a Crain's New York Business panel that he is prepared to let AI handle first reads of X-rays and mammograms once regulations allow. This is the first time a leader of a major U.S. public hospital system has framed AI radiology replacement as an operational plan, not a research agenda.

A University of Cambridge study tracking more than 8,400 data centers found that surrounding land warmed by an average of 2 degrees Celsius over 20 years, with the most affected sites seeing increases of up to 9.1 degrees Celsius, impacting roughly 343 million people within a 10-kilometer radius. The study has not been peer-reviewed. As AI infrastructure scales, environmental cost is becoming measurable at the neighborhood level.

The most common silent failure in AI coding agents is not a hallucination. It is accuracy. The agent generates code that compiles cleanly and calls methods that no longer exist, because its training data is stale. SDK docs update. APIs change. Default behavior shifts. The agent has no way to know.

Here is how to fix it:

Step 1. Install Marksnip in Chrome (free, available in the Chrome Web Store). It converts any web page into a clean Markdown file in one click.

Step 2. Open the documentation page your agent needs: the SDK reference, the API guide, the setup instructions for whatever library it is working with. Click "Download" in Marksnip to export it as a Markdown file.

Step 3. Create a folder named agent-context/ at the root of your project. Drop the Markdown file into it.

Step 4. Start every agent session with this instruction: "Read the files in agent-context/ first and use them as the source of truth for this task." The context lives in the repo, not in a thread that resets after each session.

Techzip note: This works on Claude Code, Codex, Cursor, and any agent that accepts a working directory. If you are building on a fast-moving SDK, Anthropic's, Vercel's, or any other that ships breaking changes regularly, this is the simplest way to keep your agent current without touching any configuration.

The semiconductor and infrastructure layer still captures roughly 70% of total AI revenues, two years into the generative AI cycle. Agrawal maps where value is actually accruing across the stack using data from hyperscaler earnings calls and infrastructure spending numbers. The picture is more concentrated than most AI narratives suggest: building on top of AI is still far less profitable than supplying the layer it runs on.

Why it made the cut: the piece gives you the structural frame for why the infrastructure layer keeps winning, which lands directly alongside today's NVIDIA story.