The Adjacent Brief

TL;DR: Enterprise AI pilots at Starbucks, Microsoft, and Uber are failing on governance, not model capability. Separately, only 26% of companies surveyed by KPMG have comprehensive visibility into what AI is actually costing them, which helps explain why pilots stall. On the hardware side, Unitree's robotics pricing and BYD's battery persistence are drawing fresh attention to Chinese industrial strategy and what Western competitors missed.

Worth Reading

• Texas grid flags voltage risks as data centers and crypto sites exceed compliance thresholds — ERCOT's voltage compliance problem is more immediate than the long-horizon capacity debates; infrastructure risk is arriving faster than permitting cycles.
• The data center backlash: is the local opposition actually justified? — Prof G Media works through the numbers on electricity cost and grid impact — useful corrective to both the booster and NIMBY positions.
• Polymarket is funding election conspiracy content through its ad network — A prediction market built on the premise of accurate information is monetizing misinformation. The structural irony is the story.
• Doomspending is back — and the vocabulary around it is expanding — 8Ball tracks the resurgence of doom-inflected consumer vocabulary; worth reading alongside the AI liability piece for the ambient mood it captures.
• LIPS: an open-source sip-and-puff interface for accessibility input — Hackaday. Not a trend signal — just genuinely useful hardware work that deserves visibility.
• Why no new rideshare app can beat Uber and Lyft — the unit economics argument — The Rideshare Guy lays out why this is a network-effects moat problem, not a product problem. Relevant anywhere you're evaluating incumbent-challenger dynamics.
• The new interview cheat code: build them a tool before you walk in — The Landing Pad. Social proof over credentials is becoming a hiring norm, not a hack.

Connected World

BYD's battery patience was a capital allocation thesis, not just an engineering bet

BYD never abandoned lithium iron phosphate when the industry chased nickel manganese cobalt for energy density. The piece by Christopher Chico makes the competitive mechanics clear: LFP is cheaper to manufacture, safer at scale, and adequate for the majority of use cases. Western automakers optimized for the premium end and ceded the volume market. BYD optimized for volume and now commands it. The lesson is about what happens when you refuse to abandon a technology that still works for most buyers even when the press is writing obituaries for it.

Unitree's robotics pricing is a structural problem, not a PR one

China's Unitree is positioned to dominate global robotics through the same playbook BYD used in EVs: manufacture at a price point that Western competitors cannot match on current cost structures, then let volume do the work. Semianalysis makes the case that Unitree's hardware pricing reflects genuine manufacturing cost advantages that compound over time, not a temporary subsidized loss-leader. Western robotics companies are competing on capability specs. Unitree is competing on who can afford to buy and deploy at scale. These are different races, and the second one is larger.

Data centers are running into a resource constraint that can't be engineered around

Hyperscale operators are trying to fix their water consumption — because local opposition is generating permitting friction that delays builds, ahead of any regulatory mandate. Wired's reporting on the water problem lands alongside ERCOT's voltage compliance flags from data center and crypto facility connections in Texas and the electricity analysis the broader public backlash debate that Prof G Media is working through. Across all three, AI infrastructure expansion is encountering physical and political constraints that were largely absent from the capacity projections made 18 months ago. These aren't existential limits, but they extend timelines and raise costs in ways that don't show up in hyperscaler earnings calls.

Culture & Signal

The AI liability window is open — the question is who walks through it

The lawsuits building toward an AI liability reckoning — The Next Web frames the comparison explicitly: AI companies face a wave of litigation through existing legal frameworks, particularly around harm to minors, at a scale that could produce multi-billion-dollar settlement exposure. The tobacco parallel is instructive because it took decades of individual suits before the pattern cohered into something courts could act on systematically. What's different now: the litigation is starting while the products are still being rapidly deployed, not after the harm is already historical.

Europe's platform migration is a procurement story as much as a politics story

Dozens of European governments are executing migrations away from US technology platforms — Wired's catalogue is extensive: public sector contracts shifting to European cloud providers, defense ministries moving off US productivity software, universities reconsidering their Microsoft agreements. The geopolitical framing is real, but the operational driver is simpler: procurement officers have cover to change vendors that they didn't have two years ago. That cover is now institutionalized in some jurisdictions, which makes the switch stickier than a political cycle. For US platform companies, this isn't primarily a revenue story yet — but the pipeline is drying.

Polymarket's credibility problem is self-inflicted

Polymarket's ad network is running election conspiracy content from far-right influencers — Popular Information documents the specific campaigns. A platform whose entire value proposition rests on being a more accurate signal than partisan media is monetizing the least accurate signals available. Whether this is an ad ops oversight or a deliberate revenue decision matters less than the reputational exposure: any serious institutional use of Polymarket data now carries an asterisk.

The New Consumer

AI-generated creators are passing the visual test — the disclosure problem is next

AI content creators are getting harder to spot — The Verge's reporting finds that AI-generated influencer avatars are achieving behavioral and visual parity with human creators, making audience verification difficult even for attentive viewers. Dave Winer at Scripting News covers the degradation of information quality in consumer-facing digital spaces as AI-generated content proliferates. Read together, the two pieces describe the same dynamic from different angles: the volume of convincingly human-looking AI content is rising faster than any platform's ability to label it, and faster than most consumers' motivation to verify it.

The writer's voice problem has become a self-policing problem

The piece from Every — My Editor Caught Me Sounding Like AI. Now AI Catches Me First. — makes the point: writers are pre-emptively running their own drafts through AI detectors before editors do, which means the editing relationship has acquired a new layer of self-surveillance. The irony is that the tools writers are using to avoid sounding like AI are the same tools trained on AI output. The feedback loop doesn't resolve the authenticity question; it just moves it earlier in the process.

Brand & Growth

Nothing Phone's growth is an earned media case study, not an influencer spend story

TikTokers switching from iPhones to Nothing Phones — Android Central frames this as a style and identity story, which it partly is. But the mechanism underneath is more useful: Nothing is generating organic creator content without paying for it, because the phone is visually distinctive enough to be worth filming. The glyph interface, the transparent back, the aesthetic is doing the distribution work. For brand strategists, the question this raises is "have we built something that someone would want to film?"" Most products haven't.

The demo-as-application is becoming a hiring norm

Building a tool for your interviewer before the meeting — The Landing Pad documents a pattern where candidates show up with working software built for the specific company, not a résumé. This is partly an AI capability story — building a functional prototype in a weekend is now achievable for non-engineers — but more importantly it reflects what hiring managers are actually rewarding. Credentials are a proxy for capability. A working demo is the capability. When the proxy and the thing diverge, the proxy loses.

Commerce Rewired

CFOs are flying blind on AI costs — and they know it

Only 26% of companies have comprehensive visibility into their AI costs (paywall), per KPMG's survey reported in the Wall Street Journal. Half have partial visibility, 22% have none or only see costs after billing. This is a governance problem, not a technology problem — most major cloud providers surface usage data. No one owns the number. AI spend is distributed across product, engineering, and operations budgets without a central line item, which means no one is accountable for optimizing it either. For AI vendors, this is a sales environment where the buyer doesn't know what they're already spending; for CFOs, it's a liability.

The rideshare market is a sealed room

Why new rideshare apps keep failing — The Rideshare Guy's analysis comes down to a straightforward unit economics problem: Uber and Lyft have driver supply and rider density that new entrants cannot replicate without burning cash at rates that no VC will sustain post-2022. Better product, nobler mission, fairer pricing — none of it matters if you can't get a car in under four minutes. This dynamic recurs in platform markets with two-sided network effects and density requirements. The lesson is not specific to rideshare.

Machines & Minds

Apple made Siri do something ChatGPT can't

Apple fixed Siri — David Pogue's review is specific about the capability: Siri can now autonomously open and interact with Mail, Messages, and Maps without user input, operating across native applications in a way that ChatGPT's current architecture doesn't support. This is a narrow but meaningful distinction. Apple's advantage here is platform access, not model quality. Deep OS integration lets Siri do things a third-party model literally cannot do, regardless of how capable that model is. For the enterprise AI conversation, this matters: the value of the model increasingly depends on what it can touch, not just what it can say.

The tools finding vulnerabilities are the same tools adversaries want

The AI models discovering 10,000 vulnerabilities are the same ones China is trying to copy — The Next Web reports that Google's AI vulnerability detection system was specifically targeted by Chinese espionage actors after it discovered a zero-day exploit. The dual-use problem here is unusually precise: the same capability that makes defensive AI valuable makes it a high-priority acquisition target. This is distinct from the general "AI is dangerous" framing — the specific attack surface is the most capable defensive tools, which means the organizations best positioned to protect against AI-enabled attacks are also the most exposed to having their tools stolen.

The infrastructure layer fight is getting concrete

Snowflake and Databricks are competing for the agentic AI back end — SiliconAngle's analysis frames this as a fight over who sits between the enterprise data layer and the model layer as agentic workflows become standard. Both companies are moving toward the same position from different directions: Snowflake from the data warehouse side, Databricks from the data engineering side. The model makers — Anthropic, OpenAI, Google — want that position too. Who controls the agentic client determines who captures the recurring value once the pilot phase ends and production deployments begin. That's the actual prize, and the competition for it is just getting organized.