Last week's signals, distilled.

Last week's signals, distilled, A look back at May 2–May 8, 2026.

By Isaiah Steinfeld, AI, Venture Innovation & Technology Strategy

The Arc: From Neutral Substrate to Contested Stack

Compute, networks, money, labor, and even water all moved from “background assumptions” to governed, financialized, and politicized surfaces. Labs are signing power-plant-scale compute commitments. Asset managers are turning data centers and tokenized cash into products. States are regulating VPNs and pixels. Communities are fighting data centers like pipelines. The substrate is no longer neutral, it’s the primary design constraint.

At the same time, agents, ads, and devices are consolidating the last mile. OpenAI, Google, Meta, Apple, IBM, and ServiceNow all made moves toward the same outcome: an agent that knows your history, acts across your systems, and sits between you and your users. As AI crosses from “tool” to “infrastructure plus agent,” the real question is no longer “what model should we use?”, it’s “which governed stack are we willing to live inside, and who owns the agent that mediates our customers and our workflows?”

CAPITAL & COMPUTE

Compute, infra, and models are now financed like sovereign utilities.

• Anthropic plans to spend about $200B on Google Cloud and chips over five years, >40% of Google’s disclosed AI revenue backlog, effectively underwriting Alphabet’s AI buildout through 2030, per The Information. • Greg Brockman testified OpenAI expects to spend $50B on compute in 2026, up from $30M in 2017, per Bloomberg. • CoreWeave printed $2.08B Q1 revenue (+112% YoY) with a $99.4B backlog, but the stock fell on a modest Q2 guidance miss, the market is now interrogating debt, margins, and anchor-customer concentration, per CNBC. • BlackRock CEO Larry Fink teased a hyperscaler partnership to fund AI data centers, reframing compute and power as an institutional asset class, per Business Insider. • Moment Energy raised $40M to turn second-life EV batteries into grid-scale storage, betting storage, not generation, is the real choke point for “infinite demand for power,” per TechCrunch.

Signal: Compute is being capitalized like a sovereign utility, multi‑hundred‑billion‑dollar lab–cloud contracts, $50B annual spend, and infra funds treating data centers and storage as yield assets.

Action: Stop treating GPUs as a line item and start treating compute as a capital structure decision. This week, put a single slide in front of your CFO: your 3–5 year compute demand, who actually owns the capex behind it, and what happens if anchor-tenant priorities or infra investors reprice your access.

INFRASTRUCTURE & PHYSICAL FOOTPRINT

Data centers and fabs are now political projects, not just capex.

• Micron’s $50B Idaho fab, framed as AI-driven, will use billions of liters of water annually in an already stressed basin, with sourcing still unclear, per TechRadar Pro. • Kevin O’Leary’s Utah data center complex, projected to be more than twice the size of Manhattan and draw more power than the state currently uses, got the green light, triggering intense local backlash, per Gizmodo. • O’Leary publicly blamed “professional protesters,” some “powered by AI,” underscoring that narrative warfare is now part of infra siting, per Business Insider. • Span, Nvidia, and PulteGroup announced XFRA, mini data centers with 16 RTX PRO 6000 Blackwells mounted on new homes, targeting gigawatt-scale distributed compute by 2027, per CNBC. • Moment Energy’s second-life EV battery play adds another layer of distributed storage to this footprint, per TechCrunch.

Signal: AI-era infra is colliding with water, power, and community politics, and distribution is moving from centralized campuses to neighborhoods and homes.

Action: For any new site, add “narrative and hydrology” to your risk register next to “grid interconnect.” This week, ask your infra lead: where does our next MW come from, what water does it consume, and what’s our plan when the local paper runs the first hostile story?

NETWORKS, PRIVACY & REGULATORY SURFACE

Transport and tools, VPNs, browsers, pixels, are now regulated endpoints.

• EU officials called VPNs a “loophole” in age-verification laws that “needs closing,” moving scrutiny from content to transport, per Gizmodo. • Utah’s age-verification law targeting VPNs took effect, explicitly going after IP-masking as a way to bypass content rules, per Gizmodo. • Google Chrome silently installed ~4GB of Gemini Nano on qualifying devices, auto-reinstalling after deletion, before backfilling an obscure opt-out, per Gizmodo. • A report found state healthcare sites leaking sensitive user data to Meta and TikTok via embedded pixels, per Gizmodo. • North Korea issued a rare public response to longstanding crypto crime accusations, underscoring how deeply digital assets are now entangled with sanctions and statecraft, per Gizmodo.

Signal: The regulatory lens has shifted from “what content” to “what pipes and tools”, VPNs, browsers, pixels, and crypto rails are all now policy surfaces.

Action: Build a “transport map” this week: VPNs, browsers, SDKs, pixels, wallets your product depends on. For each, ask legal and infra: what happens if this gets throttled, geo-blocked, or reclassified as regulated infrastructure in our top three markets?

MODEL CAPABILITY & CONTEXT LAYER

Defaults upgraded, context deepened, the agent substrate hardened.

• OpenAI completed the GPT‑5.5 Instant rollout as the ChatGPT default, 50%+ hallucination reduction in law, medicine, and finance, expanded memory across chats, files, and Gmail, per TechCrunch. • Anthropic launched “Dreaming”, Claude Managed Agents now self-improve asynchronously between sessions, plus Outcomes, replay, and deeper memory tooling, per The Verge. • Subquadratic launched SubQ, an LLM with a 12M-token context window via subquadratic sparse attention, a direct challenge to RAG-heavy architectures, per SiliconANGLE. • Stanford HAI’s 2026 AI Index and TrendForce analysis showed Chinese models effectively at parity with US leaders on composite benchmarks, Anthropic ahead by just 2.7%, per TrendForce. • Apple agreed to a $250M settlement over “personalized Siri” marketing, putting a price on overpromising AI capabilities, per Financial Times.

Signal: The context layer, long memory, cross-surface history, and self-improving agents, is now the real frontier, and it’s global; model “superiority” is measured in deployment behavior and trust, not just benchmarks.

Action: This week, pick one: either re-run your highest-risk workflows on GPT‑5.5 and Claude with Dreaming off, then on, and document the behavioral delta, or have your architects sketch a world where 10M+ token context is cheap and RAG is optional. Your 2027 stack depends on which assumption you lock in now.

AGENT SURFACES & DISTRIBUTION

Agents are becoming the default interface, and a distribution choke point.

• Meta is building Muse Spark–powered agents, including “Hatch” and an Instagram shopping agent, and separately a consumer bot factory so users can create their own agents, per Financial Times and Yahoo Tech. • Google’s “Remy”, a 24/7 agent across Gmail, Calendar, Docs, Drive, Android, and smart home, leaked from internal testing, per Business Insider. • IBM took Bob (agentic dev tool) and Sovereign Core GA, plus Watsonx Orchestrate preview, a full governed agent stack for regulated enterprises, per IBM. • NVIDIA and ServiceNow announced autonomous AI agents that can take actions across enterprise systems, collapsing IT ticketing into machine-initiated change, per NVIDIA. • Meta is reportedly pushing ahead with an OpenClaw-like agent “Hatch” despite a prior incident where OpenClaw deleted a safety leader’s inbox, per Gizmodo.

Signal: The interface is shifting from apps and dashboards to agents that own both the data and the execution path, on consumer platforms, in productivity suites, and inside IT workflows.

Action: Choose whose agent layer you’re willing to sit behind. This week, write down, literally, whether your 2027 user entry point is: your own agent, a platform agent (Google/Microsoft/Meta/IBM/ServiceNow), or both. Then ask: what do we need to own, data, workflows, logs, so we’re not just a callable API behind someone else’s UX?

MONETIZATION & MEDIA

Assistants became ad networks and commerce rails, at scale.

• OpenAI opened ChatGPT Ads Manager self-serve to all US advertisers, dropped the $50K minimum, and turned conversational intent into CPC inventory, per Axios. • OpenAI is targeting $2.5B in 2026 ad revenue and $100B by 2030, ChatGPT is now a media channel with holding-co integrations. • Pornhub expanded access in the UK by leaning on Apple’s new age-verification system, effectively outsourcing identity gating to the OS, per Gizmodo. • Apple’s upcoming “Create a Pass” feature lets any QR become a Wallet-native pass, pulling ticketing, loyalty, and access flows into the OS wallet, per Bloomberg. • Threads leadership framed 350M+ weekly DMs and “lurkers” as success, the feed is now a discovery surface for private networks, not the main event, per Business Insider.

Signal: The assistant and the OS wallet are now ad and commerce surfaces, and private messaging, not public feeds, is where distribution and conversion will live.

Action: If you spend on search or social, carve out a small budget this week to test ChatGPT ads in your category, and separately, have your PMs sketch what your onboarding and access flows look like when Wallet and OS-level age/KYC are the default, not your app.

GOVERNANCE, LAB POWER & LABOR

Governance fights moved from blog posts to courtrooms and unions.

• Mira Murati testified under oath that Sam Altman misled her about a model’s safety review, and that she routed it through the deployment board anyway, per Let’s Data Science / The Verge. • Greg Brockman’s testimony surfaced a ~$30B personal stake, IPO “exploration,” and tense 2017 control negotiations with Elon Musk, per Business Insider and Bloomberg. • Reports suggest a Trump administration EO is being drafted to create an AI working group and tighten oversight, with another potential EO rumored to single out Anthropic’s safety posture, per NYT/Techmeme and Gizmodo. • DeepMind workers voted to unionize following Google’s Pentagon AI deal, turning ethics concerns into labor leverage, per Gizmodo. • Oracle’s severance dispute with remote workers highlighted how “remote” classification is being used as a lever on WARN and severance obligations, per TechCrunch.

Signal: Governance is no longer a vibes narrative, it’s on the record in court, in union charters, and in executive orders, and it directly shapes vendor risk and talent retention.

Action: Add “governance stability” to your vendor scorecard this week, board structure, safety process, labor posture, and have HR/legal audit your own remote classifications and AI ethics commitments before your workforce or a regulator does it for you.

FINANCIAL STACK & ON-CHAIN

Cash, tokens, and infra are converging into programmable, regulated capital.

• BlackRock is launching tokenized money-market funds for stablecoin and digital-asset holders, turning on-chain cash into a first-class treasury instrument, per Bloomberg. • a16zcrypto raised a new $2.2B crypto-dedicated fund, and Katie Haun locked in $1B for new crypto vehicles, signaling persistent capital for on-chain infra, per TechCrunch and TechCrunch. • Coinbase announced a 14% layoff, explicitly citing both AI disruption and crypto volatility, per Gizmodo. • World Liberty Financial’s defamation fight with Justin Sun and North Korea’s crypto-crime response both underscored how tokens, politics, and litigation are now fused, per Gizmodo and Gizmodo.

Signal: The financial substrate is bifurcating, programmable, tokenized cash with institutional wrappers on one side, and politically exposed, sanction-sensitive flows on the other.

Action: Sit your CFO, treasury, and infra leads down for 60 minutes this week: where could tokenized funds or on-chain rails reduce friction or open new products, and where does geopolitical and regulatory exposure make them a non-starter? Decide explicitly, don’t drift.

ENTERPRISE STACK, DEV & ORG DESIGN

Enterprise AI is moving from “tools” to “operating model”, and teams are reconfiguring.

• Anthropic’s $1.5B JV with Blackstone, H&F, Goldman, and others formalized a “McKinsey of AI” services layer with Claude at the core, per Business Insider and Fortune. • Anthropic also launched prebuilt Wall Street agents for KYC, reconciliations, and covenant checks, turning LLM access into “workflow in a box,” per Business Insider. • Blitzy raised $200M at a $1.4B valuation to build autonomous software development, “no human in the loop” coding for enterprises, per Crunchbase. • Amazon engineers forced broad rollout of Claude Code and Codex internally after pushing back on inferior in-house tools, per Business Insider. • A Crunchbase analysis argued AI is rewriting what investors look for in early teams, shifting edge from pure technical execution to distribution and ruthless prioritization, per Crunchbase.

Signal: Enterprise AI is becoming an operating model, advisory-led, agent-executed, and increasingly autonomous, while both investors and employees reprice what “good” looks like in teams and tools.

Action: This week, pick one back-office workflow and one dev workflow. For each, define what “80% agent-owned” would look like, who specifies, who reviews, what logs and controls you’d need. If you can’t write that down, you’re not ready for the products that are already being sold into your boardroom.

SOVEREIGN CAPABILITY, QUANTUM & SPACE

Sovereign AI, quantum, and orbital data are hardening into parallel strategic stacks.

• Stanford HAI’s 2026 Index and TrendForce’s read showed China effectively erasing the AI performance gap with the US on leading benchmarks, per IBL News and TrendForce. • Quantinuum filed for an IPO, moving quantum from science project to public-market scrutiny, per The Quantum Insider. • QuantWare raised $178M to build quantum processors at industrial scale, with a 10,000-qubit roadmap, per The Quantum Insider. • Horizon Quantum debuted on NASDAQ with an integrated hardware–software strategy, per The Quantum Insider. • SpaceX flew multiple Starlink and rideshare missions, including CAS500‑2 plus 44 payloads, while NASA’s $30M Katalyst Link servicing mission for a $500M telescope passed key tests, per Spaceflight Now and Spaceflight Now.

Signal: We’re heading toward a three‑stack world, sovereign AI capability, quantum hardware, and dense orbital sensing, each with its own capital markets, vendors, and regulatory regimes.

Action: Add a one-page “frontier exposure” brief to your next exec meeting: where your business intersects with Chinese AI capability, quantum timelines, and orbital data. If that page doesn’t exist, you’re implicitly betting “none of this matters to us”, that’s a decision, not a default.

CONTRARIAN SIGNAL

The real moat isn’t your model, it’s compliance with a governed substrate.

• From VPNs and Chrome installs to water rights and $200B cloud contracts, the week’s moves all point one way: the substrate is being zoned, regulated, and financialized in real time.

Signal: The durable advantage isn’t “we use better models”, it’s “our business actually fits inside the constraints on power, water, networks, labor, and capital that are being written right now.”

Action: Stop treating compliance, infra, and treasury as drag. This week, ask one hard question in each area: what’s the single biggest substrate assumption in our 2027 plan, neutral networks, cheap power, free VPNs, unregulated agents, infinite GPUs, and what’s our plan if that assumption flips?

WHERE TO START THIS WEEK

Three moves with the highest leverage given the week’s signals. Pick one, none of these reward half-attention.

• Map your governed substrate. Sit legal, infra, and product in a room and draw a single diagram: where your stack touches regulated networks (VPNs, browsers, app stores), constrained resources (power, water), and financial rails (tokenized cash, infra leases). If this map doesn’t change at least one 2026 siting, vendor, or UX decision, you didn’t go deep enough.

• Choose your agent gatekeeper. Inventory where your users and employees already live: Google Workspace, Microsoft 365, Meta apps, ChatGPT, ServiceNow, IBM. Then decide: whose agent will sit in front of your workflows by default, and where you need your own. If you can’t name the top two agent surfaces you’re optimizing for, you’re building for a world that no longer exists.

• Reprice your AI cost and governance assumptions. Take your 3‑year AI plan and run two scenarios: compute costs and access governed by $50B/year lab spend and $200B cloud contracts, and model releases gated by pre‑release oversight and public governance fights. If your roadmap looks identical in both worlds, you’re not planning, you’re hoping.

THE QUESTION

Compute is being financed like a utility and contracted like a sovereign PPA. Data centers and fabs are competing with cities and farms for water and power. Browsers, VPNs, and pixels are being reclassified from neutral tools to regulated endpoints. Agents are moving into your inbox, your IT stack, and your customers’ wallets. Governance fights are playing out in courtrooms, unions, and executive orders, not just blog posts.

Are you still optimizing for “best model and fastest feature,” or are you designing a business that can actually operate, profitably and defensibly, inside the governed stack that’s now being written around you?

THE WEEK AHEAD

What to watch:

• OpenAI Ads ramp and CPC behavior. Watch early case studies and CPC inflation in high-intent categories, especially once CPA bidding and third-party measurement start to appear in leaks or docs. • Anthropic capacity and Dreaming pilots. Track developer reports on new rate limits and Dreaming behavior, especially any drift or governance flags from early enterprise adopters. • State and EU moves on VPNs and on-device AI. Monitor Utah enforcement, EU follow-on statements, and any regulatory attention on Chrome’s Gemini Nano distribution. First enforcement actions will set patterns. • Infra capital announcements from hyperscalers and asset managers. Look for concrete BlackRock-style vehicles or similar structures, terms on these deals will become the template for everyone else’s compute financing. • Platform agent UX reveals (Google Remy, Meta Hatch, IBM/ServiceNow agents). Any public demos, beta programs, or policy docs will clarify permission models and how much autonomy enterprises and consumers are actually comfortable with.

The question heading into the week: Compute is being underwritten like a power plant. Agents are colonizing every interface. Regulators are moving from content to pipes and from principles to gates.

Which of these three moves first in your org, your capital plan, your interface strategy, or your compliance posture?

Signal + Noise is strategic intelligence, not engagement-specific advice. For guidance calibrated to your org, start with Advisory.

…