Theses›Datacentre & AI compute

Datacentre & AI compute

What the AI buildout buys: the chips, memory, wiring, power and cooling inside the datacentre, and the software that runs on them.

  • 49 theses
  • 7 bull
  • 28 bear
Early1543359464621Open41Reopening7Closing3414222444321642298332637272311118282031251213Shut394517364019538210330Not dated474849 Early1543359464621Open41Reopening7Closing3414222444321642298332637272311118282031251213Shut394517364019538210330Not dated474849
Fig. 1Where the bets sit: windows from Early to Shut, left to right; each plot's area is the market's size by 2035. A double outline is a bull call, a single one a bear call, a dashed one a thesis with no call written yet. Washed: Beyond-CMOS devices, the largest bull call. 4 not yet sized are drawn small.
The marketsnumbered as the plan
  1. 1Datacentre thermal materials
    • Closing
    • $19B
    • Support 94
     AI accelerator thermal management does not converge on a single winning material through 2030.
  2. 2Dielectric coolants
    • Bear
    • Shut
    • $2.5B
    • Support 94
     The 3M PFAS exit dislocates the datacentre dielectric-coolant supply
  3. 3Liquid metal TIM
    • Bear
    • Shut
    • $538M
    • Support 94
     Liquid-metal TIM becomes the default die-level interface on >1kW accelerators
  4. 4In-package cooling
    • Early
    • $1.1B
    • Support 93
     Cooling inside the AI-accelerator package, sized once across its three routes
  5. 5Emerging non-volatile memory
    • Shut
    • $4.8B
    • Support 92
     Emerging non-volatile memory (MRAM, FeRAM, RRAM, PCM) displaces embedded flash and SRAM in specific sub-28nm sockets where flash cannot scale and SRAM leaks too much
  6. 6Probabilistic computers
    • Bear
    • Early
    • $347M
    • Support 89
     Probabilistic computers (p-bits / Ising / thermodynamic samplers) own server-side sampling + optimisation before fault-tolerant QC
  7. 7Beyond-CMOS devices
    • Bull
    • Reopening
    • $3.6B
    • Support 88
     Beyond-CMOS device layer: memory beachhead, IP-layer capture
  8. 8Datacentre digital twins
    • Bear
    • Closing
    • $36B
    • Support 87
     A runtime operational digital twin of the datacentre's compute+networking performance becomes a distinct layer (below design-time EDA, beside the orchestrators) for the non-hyperscaler builders who must model a fleet they cannot afford to over-provision
  9. 9CNT thermal interface
    • Bear
    • Early
    • $18B
    • Support 86
     Carbon nanotube thermal interface materials reach ≥$1B in annual revenue by 2030, up from <$50M in 2024, driven by AI accelerator and HPC qualification wins.
  10. 10Metal and solder TIM
    • Bear
    • Shut
    • $1.4B
    • Support 86
     The metal TIM family (solder TIM/sTIM, indium/metal foil, and phase-change metal alloy/PCMA) is the qualified, shipping volume die-level (TIM1) path for >1kW AI accelerators today, Indium-led.
  11. 11Rack power delivery silicon
    • Bear
    • Closing
    • $13B
    • Support 85
     AI rack power density is forcing a re-architecture of board- and rack-level power delivery (48V-plus distribution, vertical power delivery under the accelerator, point-of-load conversion for GPU/HBM rails), creating a power-conversion-silicon lane that is substrate-agnostic (silicon multiphase today, GaN where switching…
  12. 12Rack-scale design automation
    • Bear
    • Closing
    • $1.9B
    • Support 85
     Rack-scale multiphysics design automation captures a layer above chip-EDA by 2030, but only as an AI-native tool for the non-hyperscaler builders (neoclouds, sovereign-AI, enterprise AI factories) the incumbents underserve
  13. 13GaN power for datacentres
    • Bull
    • Closing
    • $1.1B
    • Support 85
     GaN power electronics captures 30% of AI datacentre PSU and 48V conversion by 2028
  14. 14Datacentre thermal and power
    • Closing
    • $453B
    • Support 81
     Differentiated thermal materials and power semis collect the per-rack tax; integration commoditises
  15. 15Optical DRAM
    • Early
    • $247B
    • Support 80
     Magneto-optical is the one optical bit that scales; the moat is a cornerable integrated stack
  16. 16Scale-up interconnect
    • Closing
    • $162B
    • Support 80
     The scale-up fabric is the vendor-agnostic toll - merchant silicon plus the NVLink tax
  17. 17Small models
    • Bear
    • Shut
    • $18B
    • Support 80
     Applied AI fragments from a few large frontier models toward many small, specialised, on-device models; this proliferation, not bigger models, is the binding demand driver for sub-watt edge-inference silicon
  18. 18Computational CXL memory
    • Bear
    • Closing
    • $12B
    • Support 80
     Computational CXL memory (near-data processing pushed INTO the pooled-memory tier: KV-cache management, preprocessing, vector search running on cores embedded in CXL memory devices, the XCENA and Dnotitia pattern) is a distinct lane from the CXL pooling/switching silicon already recorded as…
  19. 19Memory disaggregation and storage
    • Bear
    • Shut
    • $7.2B
    • Support 80
     CXL is a standard the incumbents wrote, so the fabric closes by acquisition to Astera, Marvell and the memory makers; controllers are taken, and the layer underneath is the one to watch.
  20. 20Low-power edge compute
    • Bear
    • Closing
    • $7B
    • Support 80
     Device-side AI inference does not become a merchant chip category.
  21. 21Optical memory
    • Early
    • Support 80
     The restraints are winning: HBM is scarce, but standalone optical memory is a roughly $50M research market and nothing ships at volume. HBM4 and HBM5 supply should catch up around 2029. The window is early.
  22. 22Specialised AI compute
    • Closing
    • $338B
    • Support 79
     Specialised silicon fleets win; the routing and orchestration layer is the toll
  23. 23Inference infrastructure
    • Bull
    • Closing
    • $23B
    • Support 79
     Post-training becomes inseparable from inference — the data flywheel as the moat on open weights
  24. 24Datacentre scale-out
    • Closing
    • $235B
    • Support 78
     Scale-out pulls three enabling layers at once; the durable value is in the layers, not the buildout
  25. 25AI kernel synthesis
    • Bear
    • Closing
    • $3.1B
    • Support 78
     AI/search-based kernel synthesis attacks the performance-parity-kernel constraint on heterogeneous AI compute, and at expert level on novel silicon it collapses the CUDA library moat
  26. 26Photonic packaging for CPO
    • Bull
    • Closing
    • $29B
    • Support 77
     Production AI-interconnect CPO (2027-2030) is won by inorganic photonic-packaging platforms (TFLN, BTO/SiN, silicon nitride); polymer 2PP/xolography materials do not qualify
  27. 27Optical circuit switching
    • Bear
    • Closing
    • $25B
    • Support 77
     Standalone optical circuit switching fails to reach venture scale through 2030 (a zero-success base rate over 20 years, outside Nvidia/Google-scale captive buyers)
  28. 28Retrieval and memory tier
    • Bear
    • Closing
    • $8.1B
    • Support 76
     Storage value migrates to the retrieval, index and memory tier, off the commodity bytes
  29. 29Compute as an asset class
    • Bull
    • Closing
    • $70B
    • Support 72
     AI compute becomes a financialised asset class with exchange-traded derivatives and a recognised benchmark price by end-2028
  30. 30Photonic compute
    • Bear
    • Shut
    • $375M
    • Support 72
     Optical computing is a real venture-scale market, but most of it is a trap, and the selective bet is the thesis.
  31. 31In-chip microfluidic cooling
    • Bear
    • Closing
    • $3.8B
    • Support 69
     In-die microfluidic cooling moves the binding thermal layer inside the silicon
  32. 32Optical interconnect
    • Closing
    • $162B
    • Support 67
     Through 2028 the AI/datacenter optical-interconnect transition runs pluggable → LPO/LRO → CPO, not pluggable → CPO.
  33. 33Decode chips
    • Bull
    • Closing
    • $31B
    • Support 63
     Disaggregated serving is making decode, the memory-bandwidth-bound token-by-token phase, a separate silicon socket beside the GPU rack (NVIDIA Groq 3 LPX next to Vera Rubin, d-Matrix Raptor in MGX trays).
  34. 34Data infrastructure
    • Bear
    • Closing
    • $505B
    • Support 62
     Durable value in the AI data stack accrues to whoever owns the WRITE path and the data gravity — storage, governance, pipelines, streaming, observability — not the read/retrieval path; retrieval commoditises into Postgres extensions and model-API context windows while lakehouse…
  35. 35RF waveguide interconnect
    • Bear
    • Early
    • $23B
    • Support 62
     RF-over-dielectric-waveguide is a funded third physical medium for AI scale-up interconnect, claiming the mid-range (roughly 1-10m) socket at 800G-plus lane rates with a purely electrical front-end.
  36. 36Agentic compute silicon
    • Bear
    • Shut
    • $16B
    • Support 62
     The shift to fleets of autonomous agents moves the binding compute bottleneck off the GPU forward-pass and onto orchestration, tool-execution and retrieval silicon by 2026
  37. 37AI data supply
    • Bear
    • Closing
    • $28B
    • Support 60
     AI data supply — the coverage-bottleneck stack (map/spine)
  38. 38GaN-on-si substrates
    • Bear
    • Shut
    • $2.5B
    • Support 60
     Merchant GaN-on-silicon power epitaxy leaks to the device makers who grow their own; only vertical GaN and defence GaN-on-SiC keep a door open.
  39. 39AI accelerator silicon
    • Bear
    • Shut
    • $1163B
    • Support 58
     General-purpose AI accelerator silicon (a new GPU, or a horizontal training/inference ASIC competing with NVIDIA / Google TPU / AMD / the well-funded ASIC challengers) is a real, enormous, fast-growing market whose durable value is captured by listed incumbents (NVIDIA…
  40. 40Photodetection
    • Bear
    • Shut
    • $9.9B
    • Support 58
     Datacom photodetection (Ge-on-Si photodiodes, III-V APDs, the receiver/TIA front-end) is real and necessary but not standalone-investible.
  41. 41Photonic light sources
    • Bear
    • Open
    • $72B
    • Support 57
     Across the whole photonics stack, the layer that actually captures value is the LIGHT SOURCE (lasers-on-silicon, frequency combs, external laser modules), not the modulator or the passive PIC.
  42. 42Hyperscaler ASICs
    • Bull
    • Closing
    • $84B
    • Support 54
     As hyperscalers in-source AI silicon (Google TPU, AWS Trainium/Inferentia, Microsoft Maia, Meta MTIA), the AI-compute profit pool partially migrates off the merchant GPU.
  43. 43Decentralised AI
    • Bear
    • Early
    • $35B
    • Support 51
     By 2030, a top-3 large language model by number of users will be a decentralised AI.
  44. 44Fog computing middleware
    • Bear
    • Closing
    • $187B
    • Support 48
     As compute fragments across GPUs, custom ASICs, photonic, neuromorphic and probabilistic substrates, a middleware layer forms between applications and silicon.
  45. 45Temporal compute
    • Bear
    • Shut
    • $20B
    • Support 48
     Time-domain compute stays a lab technique: no temporal-logic design ships a qualified production part or exits above $100M by 2031 (watch)
  46. 46Converter plane cooling
    • Early
    • $7.6B
     The power-converter plane is a second thermal path with no cold plate on it
  47. 47AI thermal substrates >1000 W/mK
    • Not dated
     AI compute is increasingly interconnect/thermal-bound, not purely transistor-bound.
  48. 48AI-hardware circular economy
    • Not dated
     AI-hardware circular economy: GPU reuse + critical-metal recovery
  49. 49ML closes the signal-to-noise gap on non-invasive BCI
    • Not dated
     The standing bottleneck on non-invasive brain-computer interfaces is signal-to-noise: scalp EEG, MEG, fNIRS and functional ultrasound are noisy, low-spatial-resolution proxies for neural activity, which is why useful BCI has needed surgery.