Last week's signals, distilled.

Last week's signals, distilled, A look back at Feb 28–Mar 6, 2026.

By Isaiah Steinfeld, AI, Venture Innovation & Technology Strategy

The Arc: From “AI strategy” to dependency management

Defense, infra, and hospitality all told the same story this week: AI is no longer a product decision, it’s a dependency decision. Defense ministries are treating labs as classified infrastructure. Second‑tier fabs are raising prices into a seller’s market. Restaurants are rebuilding around software and agents as the primary operating system. Capital is rotating into hard assets, compute, launch, energy, autonomy, while public markets punish anything that looks like a thin, replaceable SaaS layer.

The reframe: AI is now the constraint function on your business, not a feature you bolt on. Your real leverage is where you choose to hold risk: infra vs vendor, defense vs consumer, on‑prem vs hyperscaler, software vs robots. The question is no longer “What’s our AI roadmap?” It’s: “What does our dependency graph look like, and where are we over‑exposed, under‑priced, or pretending someone else owns the risk?”

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: mapping AI dependency graphs, vendors, infra, and governance, and turning them into concrete architecture and capital allocation changes for 2H 2026.

CAPITAL STACK

Capital is crowding into defense, infra, and orbital, and repricing everything else

• Anduril raised $4B at roughly a $60B valuation, treating defense as a growth‑stage asset class, per TechCrunch. • A quantum hardware player burned $216M while a rival moved toward a ~$2B public listing, public markets are now explicitly pricing deep‑tech burn vs path to liquidity. • SpaceX lined up a ~$1.75T IPO, while PLD Space raised to expand a European launch stack, orbital infra is being capitalized like core transport, not “space tech.” • Public markets sold off “AI trades” even as second‑tier fabs raised prices and boosted capex, infra is absorbing capital while application multiples compress.

Signal: The market is rewarding control over hard constraints, defense, compute, launch, and discounting anything that sits purely as a replaceable software layer on top.

Action: Reclassify your AI work: infra‑adjacent, defense‑adjacent, or thin app. If you’re in the third bucket, assume multiple compression and build a path to own a constraint, data, distribution, or hardware, not just an interface.

INFRASTRUCTURE & COMPUTE

Compute is now a geopolitical and pricing risk, not a stable input

• Asia’s smaller chipmakers joined larger peers in hiking prices as AI demand drives projected shortages, confirming a seller’s market for AI hardware. • Hon Hai reported a 21.6% year‑on‑year revenue jump on Nvidia server demand, with AI servers driving a disproportionate share of growth. • India and Oman advanced national compute infrastructure plans, treating data centers and fabs as sovereignty assets rather than pure commercial projects. • European gas prices spiked on war and flow uncertainty just as data centers and fabs scaled, energy volatility is now directly coupled to AI infra cost.

Signal: Compute is converging with energy and geopolitics, your GPU line item is now an exposure to regional conflict, fab pricing power, and national policy.

Action: Build a compute hedging strategy this quarter: multi‑region, multi‑vendor, and workload‑tiered. If your 2026 plan assumes flat GPU pricing and unconstrained power, rewrite it.

DEFENSE & CLASSIFICATION

AI labs are now treated as defense infrastructure, not just vendors

• NATO shifted its July summit agenda from tanks to drones and AI, AI is now a primary topic in alliance‑level defense planning. • A Ukrainian battlefield software company approached a $1B valuation, signaling that combat‑proven AI software is now a core asset class. • Defense ministries tightened vendor lists while AI labs clarified what “classified” actually means for model access and deployment. • Anthropic came under Pentagon scrutiny, with officials going on record about “holy cow” moments with frontier models, frontier capability is now a defense optics issue.

Signal: The line between “commercial AI” and “defense capability” is gone, labs, infra providers, and key software vendors are being pulled into national security frameworks.

Action: If you sell into or depend on defense‑adjacent labs or infra, assume export controls, classification rules, and security audits will touch your roadmap. Decide now whether you’re willing to live inside that regime or need a clean commercial lane.

VERTICAL SOFTWARE & OPERATIONS

Operators are rebuilding businesses as software‑first systems

• Multi‑unit restaurant operators are rebuilding around tech, dynamic menus, AI‑driven labor scheduling, and integrated ordering, per Business Insider. • Amazon’s “Just Walk Out” checkout struggled at retail scale, exposing brittleness when AI‑driven experiences hit real‑world edge cases. • Cursor wired agents directly into codebases and Slack, moving from “AI pair programmer” to embedded agentic workflows in the developer stack. • Intapp launched agentic AI into enterprise workflows, shifting from chat interfaces to embedded decision systems in professional services.

Signal: The real shift isn’t “AI features”, it’s operators re‑architecting core workflows around software and agents, while discovering where the rails still break under real‑world load.

Action: Pick one P&L‑critical workflow, scheduling, pricing, approvals, routing, and design it as if software and agents are the default, not humans. Then run a failure‑mode review: where does the system break in production, and who owns the override?

DEPENDENCY & RESILIENCE

Your AI roadmap is only as strong as your weakest dependency

• A viral consumer app went dark because a single cloud vendor hiccuped, one provider outage erased the entire product surface. • A robotaxi violated a basic social norm, triggering regulatory and public backlash, trust in autonomy is now a dependency, not a given. • Microsoft’s legal and procurement teams parsed “supply‑chain risk” tags in federal contracts, dependency mapping is moving into contract language. • A federal buyer pushed for “any lawful use” of AI in contracts, expanding the surface area of how your models can be used downstream.

Signal: The constraint isn’t just model capability, it’s the fragility of the stack you’ve built around it: cloud, vendors, social norms, and contract terms.

Action: Draw your dependency graph this week, cloud, models, data vendors, key APIs, and regulatory assumptions. For each, define a failure scenario and a 48‑hour mitigation plan. If you can’t do that, you’re not production‑ready.

GOVERNANCE & CONTRACTS

AI risk is moving from policy decks into binding paper

• Federal buyers demanded “any lawful use” of AI in contracts, shifting risk and flexibility into legal language rather than product docs. • Microsoft’s teams focused on “supply‑chain risk” tags, procurement is now encoding AI and infra risk into vendor selection criteria. • Defense ministries tightened vendor lists, effectively turning some AI vendors into regulated utilities for specific missions. • Anthropic disclosed IP theft attempts against its models and data, model security is now a governance and contractual issue, not just a technical one.

Signal: Governance is no longer about principles slides, it’s about who owns which risks in contracts, audits, and incident response.

Action: Sit down with legal and procurement this week. For every AI‑touching contract, ask: Who owns misuse? Who owns data leakage? Who owns model drift? If the answer is “unclear,” you’re carrying unpriced risk.

ENERGY & PHYSICAL INFRA

AI’s growth is colliding with physical constraints, power, water, and fabs

• European gas prices surged on war and flow uncertainty just as data centers and fabs scaled, AI infra is now exposed to commodity shocks. • India and Oman’s compute build‑outs are tightly coupled to power and cooling investments, data centers are being planned like industrial plants, not office IT. • Humanoid robots moved from lab demos into EV factories, Xiaomi deploying robots into production lines, turning AI into a direct consumer of industrial power and space. • Quantum hardware players burned hundreds of millions building physical stacks, capital is being sunk into long‑lead, energy‑intensive infrastructure.

Signal: AI is now a physical industry, your models run on steel, concrete, water, and gas, not just “the cloud.”

Action: If you run or depend on large‑scale AI workloads, add power, water, and permitting to your risk register. Map your data center providers’ exposure to regional energy shocks and start exploring workload placement as a lever, not an afterthought.

CONSUMER SURFACES & COMMERCE

Assistants are becoming ad networks and commerce rails

• OpenAI quietly re‑scoped commerce inside ChatGPT, turning the assistant into a transactional surface, not just an answer engine. • The “assistant stack” picked up its first real observability tax, vendors emerged to monitor and debug agentic flows as they touch payments and logistics. • Amazon’s checkout issues showed the cost of brittle AI experiences at retail scale, when the magic fails, it fails in front of customers and regulators. • Restaurants’ AI‑driven menus and labor tools are effectively turning hospitality into a software‑mediated commerce environment.

Signal: Consumer AI is shifting from engagement to monetization, assistants are now ad and checkout layers, with observability and reliability as mandatory taxes.

Action: If you’re building on assistants, decide explicitly: Are you an app inside someone else’s ad network, or are you owning the transaction? Your margin and data rights depend on that choice.

SECURITY & ESPIONAGE

Advanced models and infra are now primary espionage targets

• Anthropic exposed IP theft attempts against its models and training data, confirming that model weights and pipelines are high‑value targets. • Zero‑days scaled as a commercial service, offensive capability is being productized and sold into the same markets that are adopting AI. • Defense officials described “holy cow” moments with frontier models, raising the stakes for both offensive and defensive interest in these systems. • Quantum and AI hardware stacks burning hundreds of millions are now single points of failure, compromise here is systemic, not local.

Signal: Model and infra security has crossed into national‑security‑grade territory, if you’re running advanced models, you’re on someone’s target list.

Action: Treat your AI stack like classified infra: lock down model weights, training data, and inference endpoints. If you don’t have a named owner for AI security, you’ve already decided to accept unknown exposure.

CONTRARIAN SIGNAL

The real AI bottleneck isn’t talent or models, it’s contracts and dependencies

• Defense, federal, and enterprise buyers all shifted risk into paper, “any lawful use,” supply‑chain tags, vendor lists, classification.

Signal: The organizations moving fastest on AI are not the ones with the best models, they’re the ones that have cleaned up their dependency graphs and contract surfaces enough to deploy at scale.

Action: Before you hire another ML engineer or sign another model vendor, run a contract and dependency audit. If your legal, infra, and security posture can’t support aggressive deployment, more talent and better models just increase your blast radius.

OPERATIONALIZE THIS

• Audit: Draw your AI dependency map, models, clouds, data vendors, critical APIs, and regulatory assumptions, and flag any single points of failure. • Infra: Renegotiate at least one infra contract with explicit terms on GPU availability, failover regions, and energy‑linked pricing. • Talent: Assign a named owner for “AI security and dependency risk”, not as a committee, as a person with decision authority. • Governance: Update one key customer or vendor contract to clarify AI usage rights, data handling, and liability for misuse or drift. • Productivity: Take one core workflow, scheduling, routing, approvals, or pricing, and prototype an agentic version with clear human‑in‑the‑loop checkpoints. • Capital: Re‑bucket your 2026 AI budget into infra, application, and governance. If governance is at 0–5%, you’re under‑invested. • Resilience: Run a 2‑hour tabletop: “Cloud region down” and “Model vendor access revoked.” Document who does what in the first 48 hours. • Defense‑adjacent: If you touch defense or critical infra, start a classification readiness checklist, data segregation, access controls, audit trails.

THE QUESTION

AI is now a dependency, not a feature. Capital is rewarding those who own constraints, not those who sit on top of them. Governance is moving into contracts faster than most teams are updating their architectures. Defense, energy, and hospitality are all converging on the same playbook: software‑first, infra‑aware, risk‑priced.

Where in your stack are you still assuming stability, in vendors, energy, regulation, or contracts, that the last week has clearly withdrawn?

THE WEEK AHEAD

What to watch:

• NATO July summit agenda drafts, Look for concrete language on AI, autonomy, and vendor criteria; this will shape which commercial stacks can sell into defense. • SpaceX IPO filings and roadshow commentary, Watch how investors underwrite launch and satellite infra as AI‑adjacent transport, not just “space.” • Second‑tier fab earnings and guidance, Track further price hikes, capex plans, and any hints of allocation preferences for AI vs other sectors. • Major cloud provider status pages and SLAs, Monitor any quiet changes to uptime, regional guarantees, or GPU allocation terms. • Assistant and agent platform updates (OpenAI, Anthropic, major clouds), Look for new commerce hooks, ad formats, and observability requirements; these define the new tax on building atop assistants.

The question heading into the week: Defense is formalizing AI as classified infrastructure. Infra providers are repricing compute and power as scarce, strategic resources. Contracting bodies are encoding AI risk and flexibility into binding paper.

Which of these three moves first in your org, and are you ready for the others to follow?

…