Home›Trends 35 beliefs · 11 facts · as of 8 Nov 2026

Trends

What I believe about where technology is going, and how sure I am. Each belief's confidence is computed from dated evidence for and against it, and every thesis on this site rests on some of them.

ABCDEFGHIJKNo fact underneath182175941271121361013141516181920 ABCDEFGHIJKNo fact underneath182175941271121361013141516181920
Fig. 1Each fact is a lettered slab, and the high-weight beliefs it drives stand on it as columns, as tall as I am sure of each, hatched more densely the more theses rest on it. Beliefs no fact drives stand on their own ground. Washed: AI demand outruns expectations, the latest to move.
  1. 1Power is the binding constraint94
  2. 2Transistor scaling is over92
  3. 3Industrial policy persists86
  4. 4Deglobalisation continues85
  5. 5HBM is the AI memory bottleneck83
  6. 6The rack is the unit of compute83
  7. 7Value flees from cognition to atoms82
  8. 8Specialised chips beat general ones80

The slabs A Dennard scaling has ended · B Three firms hold leading-edge DRAM · C Minerals sit where they sit · D Learning curves hold · E Compute becomes heat · F Populations diverge by region · G Economies keep electrifying · H Populations are ageing · I Fertility is falling · J Ageing squeezes public finances · K The climate keeps warming

The beliefsby how sure I am

Near certain80 and above · 10

  1. TechnicalHigh weight Power is the binding constraint 94 Leans for By 2027 electrical power supply and thermal dissipation, not chip or memory supply, become the binding constraint on AI-datacentre scale-up, shifting the bottleneck from the fab to the grid and the cooling loop. Holds up 27 theses
  2. TechnicalHigh weight Transistor scaling is over 92↓1 Leans for Classical Moore/Dennard cost-per-transistor scaling has effectively ended, so future chip performance gains must come from advanced packaging, new materials and architecture rather than from shrinking the transistor. Holds up 20 theses · threatens 2
  3. PoliticalHigh weight Industrial policy persists 86↑1 Leans for Through 2030 governments keep subsidising sovereign semiconductor, energy and critical-materials capacity (CHIPS Acts, EU Chips Act, sovereign funds), so state money is a durable demand-underwriter, and that is what makes the sovereignty theses fundable rather than merely true. Holds up 15 theses
  4. PoliticalHigh weight Deglobalisation continues 85 Leans for Through 2030 trade, capital and technology blocs keep fragmenting and the post-1990 globalisation peak does not return, driving semiconductor decoupling, reshoring, critical-materials nationalism and divergent technical standards. Holds up 31 theses
  5. Technical Display glasses arrive 2029-2031 85 Leans for Display-less AI glasses (audio, camera, assistant, no in-lens image) remain the majority of smart-glasses unit volume every year through at least 2030, shipping roughly 13-14M units in 2026 and 27M in 2030. Holds up 7 theses
  6. TechnicalHigh weight HBM is the AI memory bottleneck 83↓1 Leans for, newest disagrees High-Bandwidth Memory stays the binding AI memory-bandwidth constraint through 2030, growing from $2B (2023) to a >=$25B baseline (bull case $150B if it displaces DDR5/6), with SK Hynix holding leadership. Holds up 13 theses · threatens 1
  7. Economic Technology turns inflationary 83 Leans for, newest disagrees AI-driven compute, memory and software pricing power has structurally reversed forty years of IT price deflation, so the US tech-price line stays non-negative through 2027 and chips become a 1970s-oil-style terms-of-trade shock. Holds up 1 thesis
  8. TechnicalHigh weight The rack is the unit of compute 83↓1 Leans for At the AI frontier the integrated rack (an NVLink-domain of GPU, CPU, DPU, interconnect and liquid cooling), not the individual chip, is the unit of design, sale and competition, because interconnect bandwidth, power density and cooling now bind harder than per-chip FLOPS, so margin and… Holds up 7 theses · threatens 6
  9. EconomicHigh weight Value flees from cognition to atoms 82↓1 Leans for As thinking approaches free, value flees from cognition to atoms — the scarce physical complements re-price up. Holds up 2 theses
  10. TechnicalHigh weightContested Specialised chips beat general ones 80 Leans for As AI workloads diverge (training vs inference, prefill vs decode, agents, edge), workload-specialised compute substrates take market share from general-purpose GPUs through 2030. Holds up 18 theses

Probable60 to 79 · 15

  1. EconomicHigh weight AI demand outruns expectations 78↑3 Leans for AI compute and inference demand keeps outrunning consensus expectations through 2030, so the build-out (compute, memory, power, thermal, photonics, packaging) never plateaus and the spend-driven book stays long. Holds up 35 theses
  2. TechnicalHigh weight Cheap cyber attacks 78 Leans for AI is making cyber attacks cheap faster than the world can patch. Holds up 7 theses
  3. Technical Quantum is roadmap-funded, not demand-funded 78↑2 Leans for Useful fault tolerance arrives 2029-2033, not before, and total quantum-computing revenue stays under $5bn through 2028. Holds up 1 thesis · threatens 4
  4. EconomicHigh weight Launch costs keep falling 76↓1 Leans for The cost to put mass in orbit keeps falling (~45x since the Shuttle era, with another ~10x in view via Starship), and that fall is the load-bearing driver under every space thesis: it grows satellite volume, makes refresh recurring, loosens mass budgets, and shifts compute… Holds up 17 theses
  5. Technical RISC-V takes a quarter of new chip designs 76↑13 Leans for, newest disagrees RISC-V, the open, royalty-free instruction set, takes at least a quarter of new chip designs by 2030, mostly in IoT, edge and AI-accelerator control, while Arm keeps phones and datacentre CPUs. Holds up 2 theses
  6. PoliticalHigh weight Chip supply splits West and China 75 Leans for Repeated US-led export controls ratchet the semiconductor supply chain into two non-interoperable stacks (a Western alliance on ASML EUV and TSMC/Samsung leading-edge, and a Chinese domestic stack on indigenised equipment), with sunk capex on both sides making reversal politically infeasible by 2030. Holds up 4 theses · threatens 1
  7. Social Public opposition caps the AI build-out 75 Leans for By 2030 social licence (community permitting, environmental opposition and the politics of labour displacement) becomes a binding constraint on the AI build-out, capping AI-capex theses regardless of how much chip and power supply exists. Holds up 2 theses
  8. TechnicalHigh weight Lithography tools become the bottleneck 74↑3 Leans for From ~2029, ASML's Low-NA EUV tool output (not packaging or memory) becomes the binding constraint on leading-edge AI-chip capacity, because ASML can build only ~100 machines a year against the ~3.5 EUV tools needed per gigawatt of compute, so the bottleneck migrates upstream from CoWoS/HBM… Holds up 4 theses · threatens 1
  9. EconomicHigh weight Models commoditise, value moves up 74 Leans for Raw LLM capability commoditises, so durable value capture moves up to the scaffolding, agent and application layer and down to chips, but because frontier labs are climbing into applications themselves the app layer is contested rather than free for independents. Holds up 12 theses
  10. Social AI augments more than it automates 71↓2 Leans for, newest disagrees Holds up no thesis yet · threatens 2
  11. Economic Robots fill the labour gap 70 Leans for Robotics (humanoids, manipulation, autonomous machines) is the decade's biggest under-priced labour-augmentation story, and the enabling component layer (actuators, sensors, power electronics, edge silicon) captures disproportionate value versus the full-stack robot makers and integrators by 2030. Holds up 14 theses
  12. Legal AI regulation doesn't bind 66↓1 Leans for Through 2030, AI/tech regulation (safety/liability, antitrust against the hyperscalers, training-data copyright) does not materially slow the frontier labs or hyperscaler capex. Holds up no thesis yet · threatens 2
  13. EconomicHigh weight Frontier training reaches $1T by 2030 66 Leans for, newest disagrees Frontier model training scales by brute force (each generation costing ~10x the last), so gen-6 needs $100B clusters drawing ~10GW by 2028 and gen-7 reaches ~$1T cumulative capex by 2030, with power (not silicon) the binding constraint. Holds up 4 theses · threatens 1
  14. EconomicHigh weight AI compute pays toll booths 64 Leans for, newest disagrees In the AI build-out the durable profit accrues to monopoly chokepoints (litho, leading-edge foundry plus advanced packaging, the per-wafer materials and consumables tax), not to the consumed commodity (memory), and gross-margin stability through the cycle is the proof. Holds up 5 theses · threatens 5
  15. Technical Bio-instruments move onto chips 61 Leans for Where the measurement transduction step reduces to a lithographically-defined structure on silicon (nanopores, ISFETs, microelectrode arrays, thin-film thermal cyclers, electrochemical and photonic transducers), the instrument built around it industrialises on semiconductor cost and volume curves; by 2030 at least three further laboratory assay categories have… Holds up 8 theses

Open questionsunder 60 · 10

  1. TechnicalHigh weight Optics displaces copper in the datacentre 58↑2 Leans for As datacentres scale out, optical interconnect displaces copper at successively shorter reaches (pluggable, then linear-drive, then co-packaged), so light wins in-datacentre data movement by 2030. Holds up 14 theses · threatens 1
  2. FinancialHigh weight AI capex stays financeable 55↑2 Split evenly Through 2030 the AI build-out stays financeable — hyperscaler cash flow, debt and private credit keep funding ~$500B-1T/yr of capex without a credit seizure or a circular-financing unwind. Holds up 4 theses
  3. FinancialHigh weight AI revenue catches up with capex 49 Leans against AI application revenue compounds fast enough that the ~8-10x capex-to-revenue gap narrows through 2030, rather than stalling into a write-down and capex-cut cycle. Holds up 6 theses
  4. TechnicalContested Physical AI deploys this decade 44 Split evenly General-purpose physical AI and robotics crosses from demos to real scaled deployment this decade, because the data and reliability bottleneck gets solved well enough for scaled use by 2030. Holds up 20 theses · threatens 1
  5. PoliticalHigh weight Europe builds a sovereign chip stack 41↑3 Leans against, newest disagrees By 2030 Europe builds defensible specialty-foundry, photonic and packaging coverage that serves its industrial customers (auto, industrial, defence, telecom) without reliance on Asian or US fabs (not matching TSMC at leading-edge), and EU semiconductor revenue grows from ~EUR44B (2023) to >=EUR100B. Holds up 6 theses
  6. Technical No single photonics material wins 34↑25 Split evenly Photonic IC platform competition (silicon, silicon nitride, InP, LiNbO3 / TFLN, BTO, polymer, GaAs) does not converge on a single winner by 2030. Holds up 3 theses
  7. Environmental Critical materials don't cap the build-out 26 Leans against Critical-materials supply (rare-earth magnets, gallium/germanium, grid copper) scales enough not to hard-cap robotics, energy or semis through 2030 — a cost/sovereignty constraint, not a volume cap, except heavy rare earths (Dy/Tb) and grid copper, which sit further upstream. Holds up 2 theses · threatens 1
  8. Social Talent doesn't bind 24 Leans against Human capital — AI researchers, fab technicians, and the skilled trades (especially electricians) for datacentre/grid construction — does not become the binding constraint that hard-caps the build-out through 2030. Holds up no thesis yet · threatens 1
  9. EnvironmentalHigh weight The grid keeps pace with AI 22 Leans against Generation, baseload (nuclear/SMR/gas) and grid/transmission build out fast enough that energy infrastructure does not hard-cap datacentre construction through 2030. Holds up 1 thesis · threatens 3
  10. Social LLMs lift productivity 20% by 2030 17 Leans against Large language models raise OECD labour productivity by 20% by 2030 versus a 2022 baseline, showing up in faster macro growth, lower prices and higher living standards. Holds up no thesis yet
The factsthe ground the beliefs stand on; they do not move
  1. ADennard scaling has ended
    • Drives 4
     The free lunch from shrinking transistors ended ~2006 and has stayed ended for 20 years. Gains now come from architecture, packaging, and specialisation, all of which cost power or money.
  2. DLearning curves hold
    • Drives 3
     Costs of manufactured technology fall predictably with cumulative production (Wright's law). The regularity is the fact; the rate for any given technology is the bet.
  3. CMinerals sit where they sit
    • Drives 3
     Where the minerals are is fixed by geology, supply is concentrated in a handful of countries, and a new mine takes ~17 years. Near-term supply is already decided.
  4. BThree firms hold leading-edge DRAM
    • Drives 3
     You cannot prompt a DRAM fab into existence — 3 makers, $15-20B and 3 years per fab, so supply cannot react to demand inside a cycle.
  5. ECompute becomes heat
    • Drives 2
     All compute becomes heat. There is no other exhaust. More compute means more power in and more heat out, by physics.
  6. GEconomies keep electrifying
    • Drives 2
     The world is electrifying — electricity's share of all energy use is rising, and the direction is locked in by capital already deployed. The pace and who supplies it are the bets.
  7. FPopulations diverge by region
    • Drives 2
     Europe and China shrink; Nigeria, Indonesia and sub-Saharan Africa grow — demographic weight shifts south.
  8. JAgeing squeezes public finances
    • Drives 1
     Fewer workers per retiree means pension and care spending eat a growing share of GDP while the tax base shrinks. The pressure is arithmetic from people already born; only its incidence is negotiable.
  9. IFertility is falling
    • Drives 1
     Fertility is falling below replacement across most of the rich and middle-income world.
  10. HPopulations are ageing
    • Drives 1
     The population is ageing — a larger over-65 share everywhere, already baked in by birth cohorts.
  11. KThe climate keeps warming
    • Drives 1
     The climate is warming and stays committed to warming for decades — direction certain, magnitude is the bet.