The AI Pragmatist — March 2026
Issue 01
Issue 01
01
You\u2019re busy. Here\u2019s the whole issue in one paragraph so you can decide whether to keep reading or forward it and move on.
02

“It’s no longer a computer. It’s a factory.” That’s Jensen Huang on a podcast describing the shift from retrieval-based to generative-based computing. The numbers behind the metaphor: Nvidia reported $192 billion in trailing twelve-month data center revenue, up 66% year over year. Huang forecasts a trillion dollars in AI infrastructure investment by 2028. Blackwell delivers 40x the performance of its predecessor. Vera Rubin, shipping H2 2026, triples that again. They need to triple their revenue year over year to hit that target. Oh boy.
If your company runs on cloud compute, your cost environment just entered a new regime. Hardware is scaling 10,000x faster than Moore’s Law, and so is demand.
03
Your direct reports are anxious. Your board wants a workforce strategy. Here\u2019s the data.

This month’s numbers
16% relative employment decline for workers ages 22–25 in AI-automated occupations since late 2022. Software dev employment in that cohort is down ~20%. No decline in AI-augmented roles. Wages are stable. The market adjusts through headcount, not pay. (Brynjolfsson, Chandar, Chen, Stanford/NBER, Nov 2025)
14% drop in job-finding rate for ages 22–25 in AI-exposed occupations. Independent confirmation from a separate dataset. Programmers top the exposure list at 74.5%. (Anthropic Nowcasting Report, March 2026)
59,000 tech layoffs in 2026 so far. Meta planning ~15,000 cuts to fund $600B in data center buildout. Atlassian cut 1,600, simultaneously hiring 800 AI roles. Microsoft froze cloud and sales hiring but kept Copilot hiring. (CNBC; The Spokesman-Review; IBTimes)
46% of workers using LLMs at work by mid-2025. Augmentation (52%) outpaces automation (45%) in actual usage patterns. Most common use: fixing bugs, at 6% of all conversations. (Anthropic Economic Index v4, Jan 2026)
What this means for your org: Two independent studies converge on the same finding: young workers (22–25) are bearing the brunt of AI displacement. The Atlassian pattern (cut 1,600, hire 800 AI roles) looks rational on a quarterly earnings call. Follow that logic forward five years.
Entry-level roles are where people learn judgment, develop institutional knowledge, and build the relationships that make them promotable. The companies cutting those roles today gain short-term margin but lose the pipeline that produces future directors and VPs. That 16% employment decline for 22–25 year-olds isn’t just a labor stat. It’s a leading indicator of a talent development crisis.
The 74.5% exposure rate for programmers hasn’t triggered mass unemployment because those roles are being reshaped, not eliminated. The distinction matters: if the job’s purpose includes AI-automatable tasks, the role gets elevated. If the job is the task, it gets displaced.
For hiring managers, the question isn’t whether to adopt AI tools. It’s whether you’re still investing in the people who’ll run the place in a decade.
04
Things your team can use today.

What it is:Enterprise agents got governed infrastructure out of the box. Nvidia’s March 16 keynote introduced the open-source Agent Toolkit with OpenShell runtime. Vera Rubin ships H2 2026 at 3.3x Blackwell Ultra performance. Adobe, Salesforce, SAP, and ServiceNow are already deploying the toolkit. (CNBC)
Who this affects: CTOs deciding between building custom orchestration vs. adopting vendor tooling. Infrastructure teams modeling 2027 capacity.
The numbers: $1T in Blackwell/Vera Rubin orders through 2027. AWS deploying 1M+ Nvidia GPUs. NVL576 racks draw 600 kW each.
What it is: The March 10 update embedded Gemini directly into Docs (one-prompt drafting), Sheets (autonomous project building), and Slides (context-aware design) as a persistent collaborator, not a sidebar add-on. (Google Blog)
Who this affects: IT leaders and procurement teams approaching an office suite renewal. If your org lives in Sheets and Docs, benchmark this before your next contract cycle.
The numbers:Not disclosed at launch. That’s worth noting. Google is betting on the experience delta, not published benchmarks.
What it is:OpenAI cut their “frontier” model costs by 90% for high-volume tasks. Mini ($0.75/1M input tokens, $4.50 output) and nano ($0.20/$1.25) launched March 17. (OpenAI)
Who this affects: Platform engineering teams managing LLM spend. If most of your inference calls are classification or extraction, nano reshapes the routing math.
The numbers: Nano output tokens cost 3.6x less than mini. Mini undercuts full GPT-5.4 by a wide margin on simpler tasks.
05
The stories that didn\u2019t make the launch threads.

What happened:On March 24, a threat actor called “TeamPCP” published backdoored versions of LiteLLM, a popular open-source library that routes requests across LLM providers, to PyPI. The malicious packages harvested credentials, SSH keys, and cloud tokens from anyone who installed them. PyPI pulled the packages within three hours.
But the attacker went further. Still controlling a hijacked maintainer account, they closed the GitHub issue reporting the vulnerability and deployed 73 compromised accounts to flood the thread with 196+ bot-generated spam comments. The project maintainers opened a second issue to coordinate a response. (Snyk; Trend Micro; Cybernews)
What it cost: Any environment running LiteLLM 1.82.7 or 1.82.8 should be treated as a full credential-exposure incident. The blast radius includes CI/CD runners, Kubernetes clusters, and container registries.
Why you should know:The supply chain compromise is familiar. The weaponized disclosure channel is not. Open-source projects rely on public issue trackers for emergency coordination. If your team depends on open-source AI libraries, the question isn’t just “are our dependencies pinned?” It’s “how would we find out if they were compromised, when the attacker can bury the warning?”
06
Where capital is moving tells you more about AI\u2019s trajectory than any product demo.

OpenAI — Add-on financing (nearing close). ~$10B, bringing recent total past $120B. Crunchbase
Shield AI — Series G + preferred equity. $2B at $12.7B valuation; acquiring Aechelon. Crunchbase
IBM → Confluent — Acquisition completed March 17. ~$11B enterprise value, $31/share cash. IBM
Rebellions — Pre-IPO round (Korean AI chip startup). $400M at ~$2.3B valuation. AI Funding Tracker
Meta — Layoffs planned (~20% of workforce). ~15,000 positions to offset $600B data center buildout. CNBC
Atlassian — Layoffs + AI hiring pivot. 1,600 cut (10%), 800 AI roles added. IBTimes
Microsoft — Hiring freeze (cloud + NA sales). Copilot team exempt, still hiring. The Spokesman-Review
Cursor — Revenue milestone. $2B ARR, doubled in 90 days, $29.3B valuation. TechCrunch
What the money is saying: Capital is concentrating at the extremes. OpenAI’s $120B+ raise and Shield AI’s $2B round are venture financings only in technical definition. At the other end, headcount is being treated as a financing lever for AI CapEx: cut roles, redirect the savings into infrastructure and AI hiring.
The IPO window stays closed. For your vendor risk calculus, that means your AI suppliers are staying private longer, burning more capital, and consolidating faster. If a critical vendor gets acquired or runs out of runway, your switching costs are the exposure.
07
These decisions will shape the market your business operates in.

The facts: On February 28, joint US and Israeli strikes on Iran triggered IRGC closure of the Strait of Hormuz. The strait carries ~20 million barrels per day of crude, a third of global seaborne oil trade.
Within two weeks, transit collapsed. VLCC charter rates hit $519,104/day. Brent crude peaked at $126/barrel. QatarEnergy declared force majeure on all LNG exports after Ras Laffan sustained direct damage (repair timeline: ~5 years). The IEA called it “the largest supply disruption in the history of the global oil market” and coordinated a 400-million-barrel emergency reserve release.
On March 26, Iran selectively reopened the strait, allowing Chinese, Russian, and Indian vessels through while blocking Western nations and charging transit tolls. No country has ever selectively opened a global shipping lane before. Iran has rejected a US ceasefire proposal; Trump extended the military deadline through April 6. (Wikipedia; CNN; Bloomberg; NPR)
Why it matters to you: Energy costs hit data center economics directly. PJM capacity auction prices reached $164.70/MW-day, with data centers accounting for $23.1 billion in system-wide costs. Equinix flagged power procurement volatility as a key risk even as its stock surged 25% on AI demand.
But one-third of global seaborne fertilizer also passes through Hormuz. US Gulf urea prices nearly doubled, from $350/ton to over $800/ton, during spring planting season. One ton of urea now costs American farmers 126 bushels of corn, up from 75 in December. Unlike oil, there is no strategic fertilizer reserve. (CNBC; Farmdoc Daily)
08
What\u2019s law, what\u2019s proposed, and what\u2019s actually being enforced.
EU: Parliamentary committees voted to extend AI Act high-risk deadlines. Standalone systems pushed to December 2027, embedded systems to August 2028. Only 8 of 27 member states have designated enforcement contacts. proposed (European Parliament)
US (federal): White House released a National AI Policy Framework on March 20, recommending federal preemption of state AI laws. proposed (WilmerHale)
US (state): Washington, Utah, Idaho, and Georgia signed chatbot disclosure and AI provenance bills into law. Tennessee passed health care-specific AI companion legislation. enacted (Troutman Pepper)
China: AI labeling rules now enforced under the revised Cybersecurity Law. Chatbots must display visible AI symbols; users must be notified they’re interacting with a bot at login and every two hours. enforced (IAPP)
09
One hyped claim, stress-tested against available evidence.

The claim:“Cursor built a proprietary AI model called Composer 2 that powers the fastest-growing AI coding tool in history.” Cursor’s March 19 announcement made no mention of an external base model.
Evidence for: Cursor fine-tuned Kimi K2.5 using reinforcement learning, training the model to behave differently on coding tasks. The product works: $2B ARR, 1M+ DAU, 50%+ Fortune 500 adoption. Cursor accessed the model through Fireworks as part of an authorized commercial partnership. (TechCrunch, March 22)
Evidence against:A developer spotted a Kimi K2.5 identifier in Cursor’s API response (kimi-k2p5-rl-0317-s515-fast) within hours of launch. Cursor acknowledged the base model only after community backlash. Co-founder Aman Sanger: “It was a miss to not mention the Kimi base in our blog from the start.” Kimi K2.5’s license requires prominent credit for products exceeding 1M monthly active users or $20M in monthly revenue. Cursor exceeds both. The base model comes from Moonshot AI (China-based), raising questions about IP provenance and data handling for enterprise customers. (VentureBeat; Benzinga)
What’s missing: Independent benchmarks comparing Composer 2 vs. raw Kimi K2.5 on coding tasks. Kimi K2.5’s full license terms. Moonshot AI’s data handling practices.
Verdict: Overhyped
The product works. The growth is real. But a $29.3B valuation was built partly on the perception of proprietary capability that doesn’t exist. If you’re evaluating Cursor for enterprise adoption, the provenance question is worth asking.
10
AI makes you more productive. Now you\u2019re more busy. Each month, one story that explores this tension.

If your team adopted an AI coding assistant this quarter, you probably noticed something odd. Tasks finish faster. The backlog doesn’t shrink.
Cursor has a million people writing code faster than ever. The most common use of Claude isn’t building new things. It’s fixing bugs (6% of all conversations). Research shows AI pair programmers accelerate task completion by 26% in field studies and up to 55% in controlled experiments. Across 269,000 participants in 11 studies, the average productivity gain is 21%.
None of this translates into less work. A 2026 Harvard Business Review analysis found that when AI saves time on routine tasks, managers raise output targets, tighten SLAs, or expand scope. Accounting teams that adopted AI saw an 18% increase in weekly client support. The freed-up hours went straight back into more work. Freelance demand for basic writing and translation dropped 20–50%, but demand for AI editors grew. The work didn’t disappear. It changed shape.
Economists have a name for this: the Jevons Paradox. In 1865, steam engines got more efficient and coal consumption tripledover 40 years. Jensen Huang described the AI version on Lex Fridman’s podcast: your AI agent finishes so fast it’s “bothering you all the time… reporting back to you. I got that done. What do you want me to do next?”
The efficiency creates the demand. The speed creates the backlog. The skill that matters now isn’t building faster. It’s deciding what not to build.
11
Worth knowing. Not worth a full section.

Artemis 2: NASA’s first crewed lunar mission since 1972 launches April 1. Four astronauts, including the first Black astronaut and first woman beyond low Earth orbit. (Space.com)
Anthropic leak: A CMS misconfiguration exposed ~3,000 internal documents on March 26, revealing an unreleased model called Claude Mythos, described internally as a “step change” with “unprecedented cybersecurity risks.” (Fortune)
Voxtral TTS: Mistral released an open-source text-to-speech model with 90ms time-to-first-audio. API pricing: $0.016 per 1,000 characters. (Mistral)
Mac Mini shortage: OpenClaw’s viral adoption turned the Mac Mini into the default always-on AI agent server. High-memory models (24GB+) are backordered 2–3 weeks; 64GB units sold out across Asia. A global consumer RAM shortage is compounding the problem. (Tom’s Hardware; TechRadar)
12

Nova Renay
Curator • LinkedIn
The events worth your time, your flight, and your out-of-office.
8 picks • San Diego • April 2026
Pitch & Grow
San Diego Bright Business Club
● Rancho Bernardo, Kiln
Practice your pitch, refine your value prop, and get a pulse on what other local businesses are doing.
Outdoors
● Torrey Pines
Movement, nature, and meaningful conversation for founders and investors in a casual setting where real relationships unfold.
Wellness
Bossity Babes (Women’s Workshop)
● Sorrento Valley, Hera Hub
Guided meditation, intentional prompts, and roundtable sharing. Understanding your inner world so you can move through business with clarity.
AI & Tech
San Diego AI Showcase #3 + Builders Roundtable
● Vista, LVLUP USA
No polished decks, no pitching—just people actively working with AI sharing what’s in progress. Early experiments, half-working ideas, honest breakdowns.
AI & Tech
Vibe Coding for Non-Technical Founders
● Kearny Mesa, Ansir Coworking
How non-technical founders can leverage modern tools and AI to build, test, and execute ideas faster than ever. Your permission to start building now.
Networking
SD Networking Events — April Business Mixer
● Little Italy
High-energy mixer. Diverse crowd of professionals, entrepreneurs, and creatives in one of San Diego’s most social neighborhoods.
Startups
Hussein’s North County SD Startup Mixer
● Leucadia, The Leucadian
North County’s startup pulse point. Less about pitching, more about presence—getting in the room and building relationships that compound over time.
Creative
● Carlsbad, Hera Hub
Free wine tasting, art shows, and networking. A relaxed, social way to connect without the pressure of a structured room.
Nova is the founder of Bossity. Subscribe to her newsletter: bossity.com

13
Everything above this is reporting. Everything below is my take.
Anthropic’s accidental leak reveals something bigger than a CMS misconfiguration. An unreleased model called Claude Mythos is being developed with language that signals a fundamental shift. Not just a better model, but what the company describes internally as a “step change” with “unprecedented cybersecurity risks.”
If that language is accurate, the way companies deploy AI changes. Today, you interact with frontier AI through an API: prompt in, response out. A model capable enough to warrant new security architecture implies a different relationship, one where the model acts, not just answers. Agents as users, not users with agents.
That means Anthropic won’t just be your model provider. It becomes your platform: new containment, new permissioning, new deployment infrastructure. A different vendor relationship and a different risk profile. If you’re making multi-year AI infrastructure bets, the question to ask your team now: are we building on a model layer or a platform layer? The answer changes your switching costs, your security posture, and your vendor dependency.
Everyone’s watching crude oil. The energy story has a floor: strategic reserves exist, alternative suppliers exist, demand destruction kicks in. That’s the headline, and the market is pricing it.
Here’s what’s underneath: one-third of global seaborne fertilizer trade passes through the Strait of Hormuz. Urea prices nearly doubled during spring planting season. Unlike oil, there is no strategic fertilizer reserve. No coordinated release mechanism. No emergency stockpile. The second-order impact on food prices and agricultural input costs may be larger and longer-lasting than the oil shock itself. I’m not sure the market is pricing that.