Digital Colliers Daily Briefing — May 25, 2026
Three threads converge today around a single question: who sets the terms of AI's next decade? In Rome, Pope Leo XIV issued his first encyclical, a 43,000-word intervention on artificial intelligence that drew unusually heavy pre-publication lobbying from Meta, Google, and Amazon. In Beijing, Huawei laid out a 2031 chip-density roadmap designed to route around US export controls. And in the supply chain underpinning all of it, new Epoch data shows high-bandwidth memory now accounts for nearly two-thirds of AI chip component costs — a structural shift that explains why every hyperscaler is scrambling for HBM allocation.
1. Pope Leo XIV's 'Magnifica humanitas' lands as a regulatory inflection point
What happened. Pope Leo XIV released Magnifica humanitas, his first major encyclical, dedicating roughly 43,000 words to artificial intelligence. According to Reuters' Joshua McElwee, the document urges governments to slow AI development and condemns "new forms of slavery" embedded in AI and tech supply chains. The New York Times reports the encyclical calls explicitly for AI regulation and for protections shielding children from hypersexualized AI-generated imagery. The Washington Post notes that a passage describing AI's unpredictability appears to reflect input from Anthropic; co-founder Christopher Olah attended the unveiling. Politico's Océane Herrero separately reports that Meta, Google, and Amazon executives met Vatican officials on April 29 as part of a quiet lobbying campaign ahead of publication.
Why it matters. Encyclicals are addressed to 1.4 billion Catholics but read by policymakers worldwide. By framing AI alongside the moral weight Francis brought to climate in Laudato si', Leo XIV is positioning the Vatican as a durable counterweight in regulatory debates that have so far been shaped largely by industry and a handful of Western capitals. The fact that three of the largest US platforms felt compelled to lobby the Holy See — a sovereign with no army and no market — is itself a measure of how much soft power the document was expected to carry.
Who is affected. Frontier labs face fresh rhetorical pressure at a moment when EU AI Act enforcement is intensifying and US state-level legislation is multiplying. Anthropic's apparent influence on the text is notable: it suggests the safety-forward camp inside the industry has found an unexpected megaphone. Catholic-aligned institutions, including universities and hospitals procuring AI tools, will now have explicit moral guidance to cite.
What to watch next. Whether bishops' conferences in the US, Brazil, and the Philippines translate the encyclical into concrete procurement or advocacy positions; whether EU officials cite Magnifica humanitas in upcoming Code of Practice debates; and how OpenAI, Meta, and Google respond publicly, having presumably failed to soften the final text.
Sources:
- In his ~43,000-word encyclical, the Pope urged governments to slow down AI development and decried "new forms of slavery" in AI and tech supply chains (Joshua McElwee/Reuters) — Techmeme
- Pope Leo XIV presents Magnifica humanitas, his encyclical on AI, calling for AI regulation, protection for children against hypersexualized AI images, and more (New York Times) — Techmeme
- Sources: Meta, Google, and Amazon execs met Vatican officials on April 29, as part of a quiet lobbying push ahead of Pope Leo XIV's first AI encyclical (Océane Herrero/Politico) — Techmeme
- A section of the Pope's encyclical describing AI's unpredictability suggests influence from Anthropic, whose co-founder Christopher Olah attended the unveiling (Washington Post) — Techmeme
2. Huawei proposes a new scaling law, targets 1.4nm-equivalent density by 2031
What happened. At an IEEE conference, Huawei proposed what it described as a new chip scaling law and committed to designing semiconductors with transistor density equivalent to a 1.4nm process node by 2031, according to Nikkei Asia. The approach emphasizes architectural and packaging gains rather than dependence on the leading-edge EUV lithography Huawei cannot legally acquire. In parallel, an analysis circulated by @bookwormengr argues that DeepSeek's recent model-level optimizations — which sharply reduce HBM requirements — could underpin a domestic Chinese stack pairing local memory, ASICs, and CPUs.
Why it matters. Two complementary efforts are now visible: Huawei attacking the hardware ceiling imposed by export controls, and DeepSeek attacking the memory bottleneck that gives Nvidia and SK hynix their leverage. Together they sketch a credible, if slower, path to a parallel Chinese AI hardware ecosystem rather than a perpetual catch-up. The 2031 date is far enough out to be ambitious but near enough that capex decisions made this year will be shaped by it.
Who is affected. TSMC, ASML, and Nvidia have the most exposure to a successful Huawei roadmap, as do Korean memory vendors should DeepSeek-style HBM-light architectures gain traction. Chinese domestic memory and ASIC vendors — CXMT, Cambricon, and a long tail of design houses — are the most direct beneficiaries. US policymakers may face renewed pressure to widen export-control scope to architectural design tooling.
What to watch next. Independent verification of Huawei's density claims; whether DeepSeek publishes the optimization details in a form other Chinese labs can adopt; and whether Washington responds with new restrictions on advanced packaging equipment, the area where China still trails most heavily.
Sources:
- At an IEEE conference, Huawei proposed a new chip scaling law and said it aims to design chips with transistor density equivalent to 1.4nm by 2031 (Nikkei Asia) — Techmeme
- A look at DeepSeek's model optimization to reduce HBM use, potentially enabling domestic memory, ASIC, and CPU makers to create a Chinese AI hardware ecosystem (@bookwormengr) — Techmeme
3. HBM hits 63% of AI chip component costs as total spend more than doubles
What happened. Epoch published an analysis estimating per-chip component costs across AI accelerators from Nvidia, AMD, Google, and Amazon, then weighting by quarterly production volume from Q1 2024 through Q4 2025. Memory's share of total component spend rose from 52% to 63% over that window. Advanced packaging (CoWoS) fell from 19% to 15%, auxiliary components from 15% to 9%, and logic dies held roughly constant near 13–14%. Total component spending on AI chips grew from about $22 billion in 2024 to $52 billion in 2025, with HBM alone accounting for roughly $20 billion of the $30 billion increase.
Why it matters. The economics of AI silicon have quietly inverted. A category most observers once treated as a supporting cost line is now the single largest input, and it is growing faster than the rest of the chip. This reframes margin analysis for Nvidia, supplier leverage for SK hynix, Samsung, and Micron, and the strategic logic behind every HBM-reduction technique — including DeepSeek's, discussed above.
Who is affected. Hyperscalers face a memory-bound cost curve that compresses the returns from any further logic-density gains. Memory vendors, particularly SK hynix, capture a structurally larger share of AI capex. Startups dependent on cloud GPU access feel the cost pass-through; alternative architectures that lean on cheaper memory tiers gain new commercial oxygen.
What to watch next. Q1 2026 HBM4 ramp pricing; whether CoWoS expansion at TSMC closes the packaging supply gap; and the extent to which model-level HBM optimizations, like those attributed to DeepSeek, propagate into production deployments at Western labs.
Sources:
Closing
The day's three stories form a coherent picture: AI's center of gravity is shifting away from pure model capability toward the constraints around it — moral, geopolitical, and physical. Pope Leo XIV's encyclical attempts to set a normative ceiling; Huawei's roadmap and DeepSeek's optimizations attempt to redraw the hardware floor; and Epoch's HBM data quantifies the supply-side bottleneck that ties the two together. The labs that thrive over the next 18 months will be those that can read all three signals at once.