The China AI Efficiency Question

The constraint economy is here. While venture capital continues flooding AI startups like Cursor at $50 billion valuations, the real story sits in the gap between funding and physical reality. Global RAM shortages may persist until 2030, creating a hard ceiling on AI scaling regardless of available capital. This forces a question the industry has avoided: what happens when you cannot simply spend your way to capability?

China's near-parity performance with American models despite spending 23 times less suggests an answer. When constraints bind, efficiency becomes the primary competitive axis. The implications extend beyond geopolitics. If hardware remains scarce and expensive, the advantage shifts from whoever can raise the most capital to whoever can extract the most capability from limited resources. This explains why government adoption patterns look increasingly chaotic. The NSA using Mythos Preview despite supply chain concerns and ministers questioning Palantir's NHS contract reflect institutions grappling with capability needs versus procurement frameworks designed for a different era.

The frontier is no longer about who has the biggest training runs. It is about who can build useful systems within actual constraints.

Deep Dive

Cursor's $50 billion valuation prices in a future that may not exist

Cursor's reported $2 billion raise at a valuation exceeding $50 billion tells you everything about how venture capital thinks about AI coding tools and nothing about whether those valuations make sense. The numbers reflect belief in massive TAM expansion, but the underlying assumption is that compute and memory remain abundant and cheap. That assumption looks increasingly wrong.

The AI coding market faces a collision between growing demand and shrinking supply. Every additional enterprise customer requires more inference capacity. Every new feature needs more RAM. Memory manufacturers expect to meet only 60% of demand by end of 2027, with shortages potentially lasting until 2030. Cursor can raise all the capital it wants. It cannot conjure DRAM chips that do not exist.

This creates a trap for late-stage AI companies. High valuations require proving you can scale revenue without proportional increases in infrastructure cost. But if memory remains scarce and expensive, gross margins compress. The unit economics that justified a $50 billion valuation at projected scale may not work at actual scale with 2027 component prices. Investors betting on Cursor are implicitly betting either that the memory shortage resolves faster than manufacturers project, or that Cursor finds dramatic efficiency gains that competitors cannot match.

For founders, the lesson is about timing. Raising at premium valuations feels like winning, but it locks in growth expectations that may be physically impossible to meet. If you are building an AI product that requires meaningful compute per user, you need a plan for a world where your infrastructure costs grow faster than your revenue. The companies that survive the next funding cycle will be those that treated hardware constraints as a design problem, not a procurement problem. Capital cannot solve shortages. Engineering might.

The RAM crisis is a product strategy crisis

The memory shortage is not a temporary supply chain hiccup. Manufacturers project meeting only 60% of demand through 2027, with some executives warning constraints could last until 2030. This is not a problem you optimize around. It is a hard constraint that forces fundamental rethinking of what products you can build and how you price them.

Companies face impossible choices. Meta already raised Quest 3 prices by $100 due to memory costs. Samsung increased prices across phones and tablets. Every hardware manufacturer that ships in volume is either raising prices, cutting specifications, or delaying launches. Software companies building AI features face the same bind. More capable models need more memory. If memory costs double, your inference economics break.

The strategic implication is that product roadmaps built on assumptions of steady hardware cost declines need revision. The past two decades trained an entire generation of product managers to assume compute gets cheaper over time. That assumption no longer holds. If you planned to add AI features to your product next year, you need to model those features with 2027 memory prices, not 2024 prices. The delta is large enough to kill unit economics.

This explains why China's efficiency gains matter more than absolute spending levels. American companies optimized for a world of abundant cheap compute. Chinese labs optimized for efficiency because they had no choice. Now that constraints bind globally, efficiency matters everywhere. The companies that assumed they could spend their way to capability will struggle. The companies that built for scarcity from day one have an advantage they did not plan for but will exploit anyway.

For hardware startups, this is an extinction-level event if you have not already locked in component supply. For software companies, it means rethinking every AI feature through the lens of inference cost per user. The winning products will be those that deliver value with minimal compute, not maximum capability.

Palantir's NHS contract reveals the true cost of vendor lock-in

UK ministers are considering breaking Palantir's £330 million NHS contract at the first opportunity, and the reason tells you everything about what happens when governments buy platforms instead of building capability. The National Health Service will have spent over £330 million and own no software, no intellectual property, and no ability to switch vendors without starting over.

This is not unique to Palantir or healthcare. It is the default outcome when organizations buy subscriptions to proprietary platforms rather than owning the systems they depend on. MPs report that only a quarter of participating NHS trusts see benefits from the Federated Data Platform, but all of them are locked in. The custom integration code that connects each trust to the platform belongs to Palantir. If the NHS leaves, it loses not just the platform but the connective tissue that makes its own systems work together.

The procurement failure here is structural, not personal. Government buyers optimize for speed and risk reduction. Platforms like Palantir offer both. The vendor takes responsibility for outcomes, handles integration complexity, and provides a single throat to choke if things go wrong. This looks like reducing risk. It actually concentrates it. When your entire data infrastructure depends on one vendor's platform and you own none of it, you have not reduced risk. You have outsourced control.

For government tech leaders and enterprise CTOs, the lesson is about what you actually buy. Subscription platforms trade long-term ownership for short-term convenience. That trade makes sense for commodity software. It makes no sense for core infrastructure. If a system is critical enough that you cannot replace it easily, you need to own it or at least own the ability to migrate. The NSA using Mythos despite supply chain concerns and UK ministers questioning Palantir despite sunk costs both point to the same problem: governments are realizing too late that platform subscriptions create dependencies they cannot afford.

Signal Shots

Blue Origin's New Glenn Stumbles at Scale : Blue Origin's third New Glenn launch successfully reused its booster but placed a customer satellite into an unusable orbit, forcing its de-orbit and destruction. The upper stage failure marks the first major setback for a system that only began flying in January 2025. This matters because Blue Origin positioned New Glenn as a key enabler for NASA's accelerated lunar timeline under the Trump administration. What to watch: how quickly Blue Origin diagnoses and fixes the upper stage issue, and whether NASA's confidence in using New Glenn for crewed lunar missions wavers. The company had considered launching its lunar lander on this mission but chose commercial payload instead.

Vercel Breach Exposes Developer Platform Risk : Hackers compromised Vercel through a third-party AI tool's OAuth access, exposing employee data and potentially customer environment variables including API keys and tokens. ShinyHunters, the group behind the Rockstar Games breach, claims responsibility. This matters because developer platforms increasingly integrate AI tools without robust supply chain vetting, creating new attack surfaces. What to watch: which AI tool was compromised and whether other platforms using similar OAuth integrations face exposure. Vercel's incident shows how AI adoption can introduce security gaps faster than security teams can assess them.

Chinese Workers Train Their Own Replacements : Tech workers in China are using tools like Colleague Skill to document their workflows for AI automation after bosses began requiring detailed process documentation. The GitHub project went viral despite being created as dark humor. This matters because it reveals the psychological cost of AI adoption when workers must actively participate in their own obsolescence. What to watch: whether this sparks broader labor organizing or policy responses around worker dignity in AI transitions. The emergence of counter-tools like "anti-distillation" software suggests employees will find ways to resist commodification of their expertise.

Solar Leads Global Energy Growth for First Time : Solar accounted for over 25% of global primary energy supply growth in 2025, the first time any modern renewable has led that metric, according to the IEA. Global electricity demand jumped 3% while overall energy demand rose just 1.3%. This matters because it marks an inflection point where clean power growth exceeds demand growth, meaning additions now displace fossil fuels rather than just meeting new demand. What to watch: whether this trend accelerates as battery storage deployment continues its rapid growth trajectory of 110 GW added in 2025. The timing is critical as AI datacenter power needs surge.

Fusion Startups Split on Revenue Strategy : TAE Technologies and General Fusion are going public well before achieving scientific breakeven, drawing criticism from investors who fear premature exposure could sour public markets on the entire sector. Neither company has demonstrated net energy gain from fusion. This matters because it reveals growing tension between companies pursuing side revenue through magnets or medical applications versus those laser-focused on power plants. What to watch: whether Commonwealth Fusion Systems uses its expected 2027 scientific breakeven achievement to go public, potentially validating the milestone-first approach. The fusion boom raised $1.6 billion in the past year but disagreement on commercialization paths could fragment the sector.

UK Parliament Probes AI Power Efficiency : British MPs launched an inquiry into neuromorphic computing and silicon photonics as potential paths to reduce datacenter energy consumption. Datacenters already use 2.5% of UK electricity with demand expected to quadruple by 2030. This matters because it signals government recognition that AI scaling and climate goals are on collision course without fundamental computing architecture changes. What to watch: whether the inquiry produces actual R&D funding or just remains exploratory. The UK faces a hard choice between supporting AI industry growth and meeting 2030 clean energy commitments.

Scanning the Wire

Palantir denounces inclusivity in mini-manifesto : The defense contractor published a screed attacking "regressive" corporate cultures, drawing scrutiny as it deepens ties with ICE and positions itself as defender of "the West." (TechCrunch)

Cerebras files for IPO amid tech offering wave : The AI chip maker's prospectus arrives as SpaceX, Anthropic and OpenAI prepare their own listings, signaling a surge of enormous public offerings after years of private market dominance. (NYT Technology)

Polymarket seeks $400M at $15B valuation : The prediction market's fundraising talks value it below rival Kalshi's $22B March valuation but up from its own $9B last October, reflecting investor uncertainty about regulatory trajectory. (The Information)

Colossal claims red wolf cloning breakthrough : The de-extinction startup says it successfully cloned endangered red wolves, though independent verification of the claim remains pending and the company has faced skepticism over previous announcements. (MIT Technology Review)

Robots set new records at Beijing half-marathon : The winning machine finished in under two hours, a massive improvement over last year's 2:40 robot time and edging closer to elite human performance benchmarks. (TechCrunch)

India threatens Apple with penalties over data refusal : Competition Commission of India scheduled a May hearing after Apple failed to submit required information following a probe that found alleged app market abuse. (Reuters)

Tinder's gender imbalance becomes CEO priority : Match Group's Spencer Rascoff says attracting women is his "primary focus" as Sensor Tower data shows 75% of Tinder users are men, contributing to declining engagement. (Financial Times)

Revolut targets US IPO in two years : Europe's most valuable startup CEO narrowed his timeline and filed for a US bank charter, marking the biggest regulatory milestone in the fintech's history. (The Next Web)

NextDC raises $1B for Sydney datacenter expansion : The Australian operator plans a 350MW facility and increased fiscal year capex guidance by $300M to $2.7B-$3B, betting on sustained AI infrastructure demand. (Reuters)

SK hynix begins production of 192GB AI memory module : The next-generation LPDDR5X DRAM is designed specifically for Nvidia's Vera Rubin architecture, addressing the memory bottleneck in inference workloads. (The Korea Herald)

Google diversifies chip partners with Marvell talks : Discussions cover a memory processing unit and inference-optimized TPU, adding a third design partner alongside Broadcom and MediaTek as Google reduces supplier concentration risk. (The Next Web)

Trump-linked datacenter project stalls as CEO exits : Fermi's Toby Neugebauer is leaving as the company struggles to secure an anchor tenant and faces construction delays on what was planned as the world's largest facility. (Axios)

Outlier

States Accelerate AI Regulation as Trump Administration Pushes Federal Preemption : The Trump administration is running a three-pronged campaign to stop state-level AI regulation: a DOJ litigation task force, Commerce Department reviews of "burdensome" state laws, and a legislative push for federal preemption. States responded by accelerating in the opposite direction, introducing 1,208 AI-related bills. This signals a fundamental tension about where technological governance happens. Federal frameworks optimize for industry velocity and national competitiveness. State laws optimize for local conditions and faster iteration on harm prevention. The collision matters because it will determine whether AI regulation follows the privacy model (fragmented state laws forcing de facto national standards) or telecommunications (federal preemption). What makes this weird is that the usual political alignments are scrambled. You have a Republican administration arguing for centralized federal control while states assert rights to experiment with local rules. The outcome shapes whether AI governance can adapt faster than the technology evolves.

The constraint economy has one advantage: it forces honesty about what we're actually building versus what we're fundraising about. See you when the chips run out.