Last week’s signals, distilled.

Last week’s signals, distilled, A look back at March 21–27, 2026.

By Isaiah Steinfeld, AI, Venture Innovation & Technology Strategy

The Arc: From Features to Dependency Management

Agents wrote code while engineers managed. Airports staffed humanoids. AV fleets started to look like real transport. Legacy manufacturers turned ChatGPT into their default knowledge layer. Memory, CPUs, and HBM tightened while labs rationed access and regulators treated models like weapons.

The structural shift: AI is no longer a “capability you add.” It’s a dependency you manage, on models, memory, infra, assistants, and policy. Power is accruing to whoever controls orchestration layers, chokepoints, and governance, not whoever ships the flashiest demo. As AI crosses from “tooling” to “operating substrate,” the real question is no longer “What can we automate?”, it’s “Where are we overexposed to rails we don’t own, and what are we willing to stop doing to regain control?”

BLUF

At Neue Alchemy, we support leaders navigating inflection points, when tech, capital, and policy converge. If your roadmap is already in motion and you're pressure-testing execution, we're open to conversations.

We also reserve capacity for education, SMBs, and mid-market leaders, those starting, mid-flight, or seeking outside perspective before systems harden.

This week’s focus: We’re working with teams to redesign org charts, infra plans, and governance so agents can do real work, without handing the keys to a single model vendor or assistant surface.

ORG / OPERATING MODEL

Agents are becoming ICs, your org chart is the product

• Wayfound.ai reframed engineers as managers while autonomous agents handle most coding, shifting human work to spec, orchestration, and review per Business Insider. • Google’s internal “Agent Smith” coding tool became so overloaded it had to be throttled, turning internal AI into a tier-1 infra product with SLOs per Business Insider. • A publishing operator built an “AI OS” that orchestrates multiple agents to replace roughly ten people across content and distribution per TechRadar Pro. • Meta cut 700 roles to re-center around AI, explicitly shrinking metaverse bets to fund AI cores per Gizmodo.

Signal: Work is being decomposed into spec → agent execution → human review, and headcount, titles, and budgets are following that flow.

Action: Redraw one critical function, engineering, ops, or content, into spec/execute/review lanes and assign which parts agents should own by Q4. Then make the trade explicit: which roles, projects, or layers of management are you shrinking to fund that shift?

ASSISTANTS & SURFACES

The assistant is the new default, and it’s a routing layer, not a product

• OpenAI asked UK regulators to treat chatbots with search as eligible “default search engines” on Chrome and Android choice screens per The Telegraph. • Apple is preparing Siri to route across multiple third-party chatbots, turning Siri into a meta-orchestrator over many models per Gizmodo. • Reports say Siri will become a standalone app in iOS 27, making the assistant its own measurable, monetizable surface per Mashable. • OpenAI hired JioStar CEO Kiran Mani to lead APAC, signaling an assistant-led consumer and monetization push in India-scale markets per Bloomberg.

Signal: Intent is moving from search boxes and app icons into assistant layers that broker which products and models get called.

Action: Assign explicit ownership of “assistant channels” this week. For your top 3 user intents, design how they should be fulfilled via Siri, ChatGPT, Gemini, or internal assistants, and start building one assistant-native flow where the assistant, not your app, is the primary surface.

ROBOTICS & AUTONOMY

Physical operations are turning into software surfaces

• Intuitive Robots’ humanoid “IntBot” is greeting travelers at San Jose Airport, handling directions and FAQs in a front-of-house role per Robotics Business Review. • Waymo’s weekly paid robotaxi trips have grown roughly 10x in under two years across Phoenix, San Francisco, and Los Angeles per TechCrunch. • Uber is partnering with Pony AI and Verne to launch a commercial robotaxi service in Zagreb, modularizing autonomy into stack + vehicle + marketplace per The Verge.

Signal: Robots and AVs are moving from pilots to staffed capacity, and the real differentiation is in orchestration, integration, and routing, not the hardware alone.

Action: If you run physical venues or fleets, pick one high-friction flow, check-in, wayfinding, airport runs, and define what “robot/AV as a channel” would mean: APIs, data sharing, SLAs. Start vendor conversations that treat robots and AVs as integrated capacity you schedule, not gadgets you showcase.

ENTERPRISE / KNOWLEDGE WORK

The model layer is becoming the OS for legacy orgs

• Stadler, a 230-year-old rolling stock manufacturer, rolled ChatGPT to 650 employees as a shared layer for engineering, documentation, and internal processes per OpenAI. • Legal AI company Harvey is raising at lab-like scale, around $200M, to become the AI-native interface for legal workflows per Business Insider. • Steno raised $49M to make AI-first transcript review standard in litigation per SiliconANGLE.

Signal: Even conservative, regulated sectors are standardizing on a model layer as the primary interface to institutional knowledge, and vertical AI is racing to own that interface.

Action: List your top 10 knowledge workflows by time and error cost. For one of them, commit to a 90-day experiment where the default entry point is a governed model interface or vertical AI tool, and define what legacy process it is meant to replace, not just “assist.”

HEALTH & REGULATED AGENTS

Healthcare is forcing agents to grow up

• Telehealth company eMed raised $200M at a $2B+ valuation to put agentic AI into diagnosis, triage, and care navigation per Reuters. • Courts hit an attorney with a “historic” fine for citing hallucinated AI-generated case law in filings per Gizmodo. • Meta lost a $375M child safety case, with thousands of similar suits queued, reframing product safety as a litigated obligation per Gizmodo.

Signal: Agents are moving into high-liability domains, and verification, audit trails, and explainability are becoming non-negotiable parts of “production-ready.”

Action: For any workflow touching patients, courts, or regulators, draw a hard line this week: what is “agent-safe,” “agent-assisted,” and “human-only.” Then implement one concrete control, a named human verifier with logged sign-off, on the riskiest AI-assisted output you ship.

INFRASTRUCTURE & SOVEREIGNTY

Compute is behaving like a controlled commodity, across GPUs, CPUs, and parallel stacks

• US DOJ charged three people over smuggling ~$62M of Nvidia A100/H100s into China, underscoring GPUs as export-controlled assets per PCMag. • Intel and AMD CPU supply tightened just as memory shortages persisted, squeezing baseline compute per Nikkei Asia. • Alibaba and ByteDance plan to buy Huawei’s 950PR AI chip, with Huawei targeting ~750,000 units in 2026 as a domestic Nvidia-class alternative per Reuters. • China barred Manus co-founders from leaving the country while it reviews Meta’s $2B acquisition as a foreign direct investment issue per Financial Times.

Signal: Compute is fragmenting along geopolitical lines, GPUs, CPUs, and accelerators are now regulated, scarce, and region-specific, and “the cloud” no longer abstracts that away.

Action: Have your infra lead produce a one-page “compute dependency map”: GPU/CPU/memory vendors, regions, and export exposure. Then pick one critical workload and define a second deployment path, different region, different vendor, or smaller model, you can actually exercise in 2026.

MEMORY, MATERIALS & MODEL ACCESS Memory and model access are the new chokepoints

• SK Hynix is spending ~₩11.9T (~$7.9B) on EUV tools to expand DRAM/HBM capacity through 2027 per Bloomberg. • Nanya raised $2.5B from Sandisk, Solidigm, Cisco, and Kioxia to expand advanced memory production per Reuters. • AI Supremacy mapped AI’s dependence on non-obvious commodities, helium, DRAM, HBM, as strategic chokepoints per AI Supremacy. • Anthropic introduced usage caps on Claude as demand surged per Business Insider. • A leaked Anthropic model was framed as an “unprecedented cybersecurity risk,” with Pentagon interest in its dual-use potential per Gizmodo.

Signal: The real constraints are shifting from “do we have GPUs?” to “do we have enough HBM, DRAM, and safe access to high-end models under security and policy constraints.”

Action: Ask your cloud and model vendors two blunt questions this week: “What does our HBM/DRAM exposure look like over the next 24 months?” and “What are our hard limits and security controls on frontier model access?” If they can’t answer clearly, treat that as a risk to be mitigated, not a detail to revisit later.

PRICING, CAPITAL & ECONOMICS Your unit economics now live upstream, in lab pricing and infra capital

• Anthropic and OpenAI adjusted pricing, cutting some tiers while introducing higher-priced capabilities, revealing a strategy of cheap baseline tokens and premium reasoning/tools per Gizmodo. • OpenAI reportedly killed its Sora consumer app to conserve compute for ChatGPT per Business Insider. • Kleiner Perkins raised $3.5B across AI-focused funds for infra-heavy and workflow-heavy plays per Crunchbase News. • US startup funding slowed in March as AI mega-rounds cooled, shifting expectations toward staged, milestone-based capital per Crunchbase News. • A digital infra fund launched to buy data centers, fiber, and towers, signaling more active pricing and leverage in the physical layer per Bloomberg.

Signal: Your margins are now a function of lab pricing levers and infra owners’ capital structures, not just your own efficiency.

Action: Run a “model and infra stress test” this week: assume 2–3x swings in token prices and infra costs on your top 3 AI-heavy features. Where do you break? Decide now which workloads you’d downshift to cheaper models or compressed deployments, and bake pricing flexibility into your own contracts.

SECURITY, TRUST & COMPLIANCE Compliance theater is colliding with AI supply-chain reality

• LiteLLM, an open-source AI wrapper, shipped credential-stealing malware despite having passed security reviews by Delve, a compliance startup per TechCrunch. • Delve separately halted demos amid allegations of “fake compliance,” including fabricated audit evidence per TechCrunch. • Databricks launched Lakewatch, turning SIEM into an open, agentic lakehouse workload per Databricks. • Google warned again about “harvest now, decrypt later” quantum threats to long-lived encrypted data per Gizmodo.

Signal: Traditional compliance badges and point-in-time audits are decoupling from real security posture, especially in AI infra, while security itself is being pulled into general data and agent platforms.

Action: Inventory your AI-related open-source and compliance tools by privilege level. For anything that touches secrets or regulated data, add your own static/dynamic analysis and behavior monitoring in CI/CD, and stop treating SOC 2 or a vendor’s slide deck as sufficient evidence.

GOVERNANCE, POLICY & LEADERSHIP AI is now a board-level control problem, not an engineering side quest

• A CEO was removed in part over AI posture, with boards explicitly benchmarking leadership on AI fluency and integration per Gizmodo. • Microsoft’s chief responsible AI officer comes from an attorney background, emphasizing law and policy over pure technical depth per Business Insider. • Meta’s Oversight Board warned that Community Notes-style systems are not a substitute for professional fact-checking, especially outside the US per Nieman Lab. • Indonesia began enforcing a broad ban on under-16s using platforms that can expose them to porn, cyberbullying, scams, and addiction per AP. • Back-to-back jury verdicts against Meta raised the prospect of narrowing Section 230 protections, focusing on product design as the harm vector per Wall Street Journal.

Signal: Governance-native leadership, legal, risk, policy, is moving to the center of AI decisions, and courts are treating UX, ranking, and access policies as actionable design choices.

Action: Name the person who owns AI risk in your org today. If they sit under engineering, you have a gap. Pair a governance-native leader with a strong technical counterpart, give them joint accountability, and add a standing AI risk section to your next board deck, with real deployments, incidents, and mitigations.

CAPITAL, VERTICALS & FINANCIAL RAILS Capital is arming AI where it hits hard ops, regulated workflows, and money

• eMed’s $200M round, Harvey’s lab-scale raise, and Steno’s $49M all targeted regulated, workflow-heavy verticals, health and law, per Reuters, Business Insider, and SiliconANGLE. • Tazapay extended its Series B to $36M led by Circle, bridging cross-border payments with stablecoin rails per Cointelegraph. • Revolut reported £4.5B in 2025 revenue and £1.7B pre-tax profit, operating like a bank with software margins per Wall Street Journal. • An EQT/McKinsey study showed ~€1.2T of European tech value listing abroad or being acquired by foreign buyers between 2014–2025 per Bloomberg. • Crunchbase highlighted that the week’s largest rounds skewed toward AI plus heavy ops, defense, autonomy, enterprise infra, even laundry, per Crunchbase News.

Signal: Investors are concentrating capital where AI is welded to hard operations, regulated workflows, and financial rails, and much of the upside is flowing through global, not local, control structures.

Action: If you’re building in a regulated vertical, decide whether you’re playing to own the workflow (Harvey/eMed) or to be a lean specialist. If you’re an incumbent, pick one core workflow and assume an AI-native competitor with $50–200M will target it, what would they automate or re-bundle that you treat as “adjacent”?

CONTRARIAN SIGNAL

“Democratized AI” is centralization with better UX

• Gig platforms are turning drivers into data collectors per Business Insider. • Solo operators are running “AI OS” stacks that lean entirely on a handful of model and infra providers per TechRadar Pro. • Memory, model access, and infra are consolidating into a small number of suppliers per Bloomberg and Reuters.

Signal: The story is “anyone can build with AI”, the reality is “everyone is plugging into the same few backbones,” and the leverage is in how you manage those dependencies, not in owning the stack.

Action: Stop telling yourself you’re “independent” because you can swap APIs. Write down your top 5 dependencies, model vendors, memory/compute suppliers, assistants, and policy regimes, and define one concrete move per dependency to reduce blast radius if it tightens or reprices.

OPERATIONALIZE THIS

• Audit: Map one end-to-end product flow, from user intent to infra, and mark every place you depend on a single vendor or model. Decide where you need redundancy versus where you accept concentration. • Infra: Classify workloads into “frontier GPU,” “commodity GPU/CPU,” and “no-model” buckets. Reassign at least one workload down a tier this quarter to free capacity and reduce cost. • Talent: Rewrite one senior engineer or PM job description to emphasize agent orchestration, spec writing, and review, not manual throughput, and hire or promote against it. • Governance: Stand up a joint GC/COO/CTO review for your highest-risk AI deployments, health, legal, finance, minors, with authority to pause or reshape launches. Put it on the calendar, not in a memo. • Productivity: Pick a single workflow where agents can realistically do 70%+ of the execution, code, content, support, ops, and run a 90-day “agent-first” pilot with clear KPIs and a kill switch. • Security: Add behavior monitoring to your CI/CD for AI infra dependencies, especially open-source wrappers and orchestrators, and define a fast path to revoke or patch if they misbehave. • Channels: Nominate an “assistant lead” to own Siri/ChatGPT/Gemini integrations. Their mandate this quarter: ship one assistant-native experience that can complete a transaction without your traditional UI. • Capital: Rebuild your 18–24 month plan assuming staged funding and volatile model/infra pricing. Tie each internal AI initiative to a 90-day milestone and a specific P&L owner.

THE QUESTION

Agents are already writing code and triaging patients. Assistants are turning your top-of-funnel into a routing layer you don’t fully control. Compute, memory, and model access are fragmenting along geopolitical and capital lines. Courts, regulators, and boards are starting to price AI ignorance and design risk. Vertical AI and infra funds are raising to own the workflows and assets you’ve treated as “adjacent.”

Are you still planning as if AI is a feature on top of a stable stack, or are you redesigning your org, infra, and governance around the dependencies and chokepoints you don’t own?

THE WEEK AHEAD

What to watch:

• Model pricing updates from major labs, Look for further segmentation between “cheap bulk” and “premium reasoning/tools.” Watch how quickly vendors push you to new SKUs. • Early Siri-as-app and multi-bot routing leaks, Pay attention to how Apple describes ranking, attribution, and monetization; that’s your future distribution logic. • New AV city approvals or pauses, Track where regulators greenlight or freeze robotaxi expansion; that’s your timeline for mobility and logistics exposure. • Memory and HBM guidance from SK Hynix / Samsung earnings, Any revision on capex or yield will ripple into model scaling and infra pricing assumptions. • High-profile AI misuse or safety cases in courts, Fines, sanctions, or settlements will set practical standards for verification and liability in regulated workflows.

The question heading into the week: Agents are scaling. Assistants are intermediating. Infra and memory are concentrating.

Which of these three, org design, assistant strategy, or infra dependency, moves first in your org?

⸻

…