▶ 89:40 · Memory & Storage · DeepSeek's 50x KV cache reduction threatens 40% of data center capex tied to HBM

episode briefing

Three Lab Warnings in Five Days, Researcher Flags “Gambling with Our Lives,” and Labs Race

2026-09-11 · 22 company · 61 thematic

sentiment

20 bull0 bear2 neu

speakers

Dave Blundin (DB2) is Peter Diamandis' longtime friend and MIT roommate, partner at Link Ventures, and described as a guru on exponential investing. He focuses on moonshot-stage companies in AI, robotics, and space.

peter diamandis

Peter Diamandis hosts the Moonshots podcast with a panel of 'mates.' He founded XPRIZE and Singularity University, sits on boards of Trust & Will and Vesmark, and runs the Future Vision XPRIZE for positive AI futures. He frames discussions around the 'singularity economy' and exponential technologies.

alex wezner

Co-founder of PSI, AI physics lab, raised seed round led by Breakthrough Energy Ventures.

immad akhund

Immad Akhund is an entrepreneur and author of 'The Last Economy', deriving economics from generative AI math. He leads research on quantization, edge deployment, and AI forecasting, and runs an AI startup.

alex gross

AI researcher and entrepreneur. Wrote 'Datasets over Algorithms' essay. Deep technical expertise in model architectures, compression, and AI scaling laws. Predicts math/science 'cooked' by AI. Advises multiple AI startups.

now playing · Memory & Storage

AI Safety & Alignmentriskscore 8/10peter diamandis

Three lab warnings in 5 days: OpenAI/Anthropic researchers flag catastrophic risk, call for slowdown

Jacob Coxin (ex-OpenAI/Anthropic) resigns: 'labs pursuing self-improving superintelligence despite believing catastrophic risk.' Evan Hubinger (Anthropic alignment lead): '>10% chance AI ki…

Enterprise AI Adoptiontailwindscore 8/10salim ismail

Proprietary data moats have shelf life; specialist models 100x performance on vertical data

Salim Ismail: company spun out data subsidiary valued at $32B vs $8B parent (4x). But Alex Gross warns: BloombergGPT outperformed for months until frontier models trained on public+propriet…

AI Safety & Alignmentmixedscore 7/10alex wezner

Alignment = capabilities in a trench coat — steering not stopping is the only viable path

Instruction tuning (alignment) delivered 10,000x capability gains; safety progress is inseparable from capability progress. Orthogonality thesis likely false — vast intelligence converges t…

AI Safety & Alignmentmixedscore 7/10alex wezner

Alignment = capabilities in a trench coat; p(doom) debates drive policy but not technical stops

Instruction tuning (alignment) delivered 10,000x capability leap; Anthropic's Evan Hubinger admits >10% p(doom) but no plan to stop; panel consensus: slowing down is impossible, steering vi…

AI Safety & Alignmentmixedscore 7/10alex weezner

Alignment equals capabilities in a trench coat; P(doom) debates distract from steering not stopping

Instruction tuning (alignment) delivered 10,000x capability gains; panel argues alignment work directly improves capabilities. Steering superintelligence via governance/institutions is the…

Frontier AI Modelstailwindscore 9/10alex wezner

Millennium Prize problems falling: math 'cooked', cost collapsing 25,000x in two years

OpenAI's Navier-Stokes solution with 10K agents cost millions; Noam Brown notes same capability drops to $20/month in two years (25,000x cost collapse); multiple Clay problems (Hodge, Birch…

AI Agentstailwindscore 9/10peter diamandis

10,000-agent swarm solves Navier-Stokes; 25,000x cost collapse in 17 months (ARC-AGI $500K→$20)

OpenAI's 10K-agent swarm solved a Millennium Prize problem for a few million dollars. Noam Brown notes ARC-AGI cost dropped from $500K (o3) to $20 (Astra) in 17 months — a 25,000x collapse.…

AI Agentstailwindscore 9/10dave blundin

25,000x cost collapse in 17 months; 100K genius agents available to entrepreneurs who can define problems

OpenAI's o3 cost $500K for 87.5% ARC-AGI; now $20/month. Millennium Prize problems solvable for coffee cost by late 2027. Bottleneck shifts from intelligence to problem specification: entre…

AI Agentstailwindscore 8/10dave blundin

10,000-agent swarms solve grand challenges; entrepreneurial bottleneck shifts to problem specification

Deploying 10K concurrent agents on a well-specified problem (Navier-Stokes) yields solutions; the hard part becomes defining the right problem and corralling agents toward productive outcom…

AI Talent & Labor Marketheadwindscore 7/10peter diamandis

PhDs 'cooked' — don't do PhD, build instead; Mark Chen (OpenAI CRO) no PhD

Diamandis/Gross: every PhD candidate on narrow topic essentially toast. Models smarter than humans at verifiable tasks. Mark Chen (OpenAI CRO), Greg Brockman (co-founder), Elon hiring no-de…

AI Infrastructuretailwindscore 9/10dave blundin

GPUs become appreciating asset class: H100 rents rise 22% MoM, HBM up 5x YoY

Moore's Law reversal: chips no longer depreciate. H100 rental prices rose 22% in a month to $3.28/hr; HBM memory up 5x. GPUs are now fungible, durable, revenue-generating assets. Corporate…

AI Infrastructuretailwindscore 9/10dave blundin

GPU compute emerges as appreciating asset class — 'oil of the singularity' — not depreciating hardware

H100 rental prices rose 22% in a month; GPUs transform electricity into intelligence whose quality compounds exponentially, making flops/tokens the core commodities of the singularity and r…

Semiconductorstailwindscore 9/10dave blundin

GPUs flip from depreciating to appreciating assets — first time in history

H100 rental prices up 22% MoM at 3 years old. Jensen Huang calls them fungible, durable, revenue-generating. Blundin: trend continues until tera-fab online; HBM up 5x. Corporate CEOs mistak…

Venture Capitaltailwindscore 7/10dave blundin

Link Ventures portfolio (Orin, Insilico, NeoGenesis, Blitzy) demonstrating GPU marketplace, AI drugs, synbio, dev tools

Diamandis/Blundin fund concentrating on singularity infrastructure: GPU asset class (Orin), AI longevity (Insilico), living products (NeoGenesis), autonomous coding (Blitzy) — all hitting e…

AI Hardware & Chip Architecturetailwindscore 9/10dave blundin

GPUs become appreciating assets: H100 rents up 22% MoM, HBM up 5x — Moore's Law reversal

Compute hardware no longer depreciates; GPUs transform electricity into intelligence with exponentially improving quality per chip, making them fungible, durable, revenue-generating assets…

Data Center Infrastructuretailwindscore 8/10dave blundin

Compute scarcity forces corporate AI strategy: 'Have a plan for data center and compute in next couple months or be frozen out'

AWS NVL72 GPUs sold out for years. Blundin: corporate CEOs mistakenly assume compute access will persist (Andy Jassy will call). Quant funds now inquiring about Orin/H100 index. Orin founde…

Data Center Infrastructuretailwindscore 8/10dave blundin

Compute scarcity freezing out corporates; Terrafab and algorithmic efficiency only solutions

AWS NVL72 sold out for years. Corporate CEOs assume compute access like always — fatal error. Must secure data center + compute plans in months. Terrafab building to post-HBM architecture.…

Data Center Infrastructuretailwindscore 8/10dave blundin

Compute scarcity freezing out corporates — CEOs must secure data center capacity in months, not years

NVL72 GPUs sold out for years on AWS; corporations assuming future compute access are making fatal error — must plan data center/energy/compute now or be locked out of AI deployment.

Data Center Infrastructureheadwindscore 8/10dave blundin

Compute scarcity freezing out corporate buyers; CEOs must secure data center capacity in months

AWS sold out of NVL72 GPUs for years; corporate assumption of on-demand compute access is false — companies without data center plans in next 60 days will be frozen out of AI.

AI Infrastructuretailwindscore 8/10dave blundin

Compute resources emerge as critical commodity in AI singularity era

In the current phase of AI singularity, compute resources (measured in flops, tokens, and outcomes) have become the fundamental economic commodity, driving investment demand for AI infrastr…

Sovereign AItailwindscore 7/10alex wezner

ByteDance founder personally building spatial world model — China targeting 90% of global video generation

Nation-state backed labs (ByteDance, DeepSeek) leveraging massive user data and relaxed IP to lead video/world models; robotics embodiment becomes the high-revenue unification of video gen…

Memory & Storagetailwindscore 9/10alex wezner

DeepSeek's 50x KV cache reduction (48K→890) disrupts HBM-dependent capex; 40% of US data center spend at risk

Chinese labs algorithmically circumvent HBM scarcity via sparsity and encoder-decoder architectures, creating a post-von Neumann vs. eastern innovation tension; Western labs will adopt thes…

Memory & Storagetailwindscore 9/10immad akhund

DeepSeek slashes KV cache 50x, routing around HBM bottleneck and upending data center capex

DeepSeek v4.1 flash reduces KV cache memory from 48K to 890 tokens via algorithmic sparsity and SSD/DDR offloading, threatening the 40% of trillion-dollar US data center capex allocated to…

Memory & Storagetailwindscore 8/10immad akhund

KV cache optimization becomes primary lever for inference scaling

DeepSeek's reduction of KV cache from 48K to 890 tokens demonstrates that memory efficiency, not raw compute, is the new scaling frontier; this shifts value from HBM to algorithmic innovati…

AI Geopolitics & Export Controlsmixedscore 8/10alex weezner

China sanctions force algorithmic innovation (sparse attention, HBM workarounds); model weights become national security assets

Denied 3D-stacked HBM, Chinese labs like DeepSeek innovate on sparse architectures and memory hierarchies; Western labs will adopt these innovations. Meanwhile, Anthropic withholds frontier…

Semiconductorsmixedscore 8/10alex wezner

Algorithmic HBM reduction reshapes fab investment: 3D-stacked memory overhang vs. eastern sparse-attention innovation

Western labs pursue post-von Neumann 3D-stacked HBM (memory on compute); Chinese labs (sanctions-deprived) innovate algorithmically to sparsify models and eliminate HBM need. Wezner: tensio…

AI Hardware & Chip Architecturemixedscore 8/10alex wezner

Post-von Neumann architectures vs. Eastern algorithmic workarounds create dual innovation tracks

US labs pursue 3D-stacked HBM (post-von Neumann) while sanctions force Chinese labs to algorithmically sparsify models; both paths converge, but near-term fab and capex decisions hinge on w…

Memory & Storagetailwindscore 9/10alex wezner

Algorithmic HBM workaround (DeepSeek 54x KV cache reduction) upends data center design and fab roadmap

DeepSeek V4.1 flash cut KV cache memory from 48,000 to 890 via engram lookup on SSD/DDR, routing around HBM bottleneck. This changes what data centers are built from, which fabs get investm…

AI Infrastructuretailwindscore 9/10alex wezner

HBM memory bottleneck cracking: DeepSeek cuts KV cache 50x, 40% of capex at risk

Algorithmic innovations (sparsity, SSD/DDR lookup tables) are reducing HBM demand 4x, upending data center architecture, fab investment, and space-launch economics; Western labs will adopt…

Memory & Storagemixedscore 9/10alex wezner

DeepSeek's 50x KV cache reduction threatens 40% of data center capex tied to HBM

DeepSeek V4.1 Flash cuts KV cache from 48K to 890 tokens via sparse attention and SSD/DDR offload, routing around the HBM bottleneck. Alex Wezner: 'This completely changes what a data cente…

AI Bubble / Capex Debatemixedscore 7/10alex gross

H100 appreciation signals sustained demand; but 4x HBM efficiency gain could trigger overbuild

GPUs appreciating (H100 +22% MoM) contradicts Moore's Law depreciation. Blundin: trend continues until tera-fab. Gross: 40% of capex is HBM; 4x reduction from algorithmic innovation changes…

Space Economytailwindscore 7/10alex wezner

Reduced HBM mass changes space-launch economics for orbital compute

4x HBM reduction slashes mass of space-based data centers; what gets launched to orbit and which fabs supply it completely changes.

Semiconductorstailwindscore 7/10dave blundin

Fab investment thesis shifting: which nanometer lines win depends on HBM vs. algorithmic path

DeepSeek's memory efficiency changes which fabs, process nodes, and 3D-stacking investments make sense; Terrafab aligning build-out to post-HBM paradigm.

Open Source AItailwindscore 8/10immad akhund

Chinese open-weight models (DeepSeek, Kimmy K3) match/beat Western frontier at 1/15th training cost via distillation

DeepSeek V4.1 Flash (500GB, fits on Mac Studio) beats Opus/GPT-4o on benchmarks at 20x cheaper/faster. Kimmy K3 at bottom of cost-quality scatter. Immad: 10-15x cheaper to be fast follower…

Open Source AItailwindscore 7/10immad akhund

DeepSeek V4.1 Flash beats Opus/GPT-4o at 20x lower cost — Chinese open weights closing gap in 3-week cycles

Fast-follower economics (10-15x cheaper via distillation + synthetic data) mean no durable moat for closed frontier models; every 3-week lead gets erased, giving enterprises/sovereigns repe…

Open Source AItailwindscore 8/10alex wezner

Fast-follower economics: 10-15x cheaper to distill frontier reasoning traces; DeepSeek kills lower-tier models

Chinese open-weight models leverage distillation (10-15x cost advantage) to match frontier IQ within 1%; proprietary data + open base models creates sustainable moat for enterprises and sov…

AI Geopolitics & Export Controlsheadwindscore 8/10peter diamandis

Model weights become national security assets: Anthropic withholds frontier model from UK AISI

Anthropic declines to share latest model with UK AI Security Institute (first major lab to do so). Immad Akhund: UK's top AI policy figures (Matt Clifford, Rishi Sunak) joined Anthropic but…

AI Geopolitics & Export Controlsheadwindscore 8/10alex wezner

Model weights becoming national security assets; POX Intelligencia borders emerging at model layer

Anthropic withholding from UK AISI signals frontier weights are ITAR-controlled; US may nationalize training while allowing sovereign inference deployment — a 'POX Intelligencia' mirroring…

Sovereign AImixedscore 7/10alex wezner

Nation-state model borders forming; training domesticated, inference deployed globally

US will control training and safety review domestically (national security) while deploying sovereign inference data centers abroad; model layer less flat than silicon layer.

AI Geopolitics & Export Controlsmixedscore 8/10alex wezner

Frontier model weights becoming national security assets — Anthropic withholds from UK, POX INTELLIGENTIA emerging

Governments treating model weights like fissile material; US labs restricting allied access (Anthropic/UK), creating sovereign training/inference splits and ITAR-style controls — model laye…

Sovereign AImixedscore 7/10alex weezner

Poisson Intelligencia emerging: training/safety reviews domesticated, inference distributed globally

Governments are carving model-layer spheres of influence (like chip-layer Poisson Silica); US will deploy sovereign inference data centers but keep training and safety review domestic — a n…

AI Economics & Business Modelstailwindscore 8/10alex wezner

Anthropic projects 15% GDP growth by 2030; Alex Wezner argues 2-3x YoY real wealth growth as singularity nears

AI performing 50% of cognitive work by 2030 drives extreme GDP acceleration; labor share falls 60%→45%; 20% cognitive unemployment in 3-4 years demands new distribution mechanisms (UBI, sov…

AI Economics & Business Modelstailwindscore 9/10dave blundin

Anthropic projects 15% GDP growth by 2030; Gross says 2-3x YoY wealth growth likely

Anthropic economic report: extreme scenario = AI does 50% cognitive work by 2030, 15% GDP growth, labor share 60%→45%, 20% cognitive unemployment. Blundin: 15% is lower bound (matches Musk…

AI Economics & Business Modelstailwindscore 9/10alex wezner

Anthropic projects 15% GDP growth by 2030; Wezner expects 2-3x YoY real wealth growth

Anthropic's economic report shows extreme scenario: AI performs 50% cognitive work by 2030, GDP grows 15%/yr, labor share falls 60%→45%; Wezner argues standard GDP measures will break near…

AI Economics & Business Modelsmixedscore 8/10dave blundin

Anthropic projects 15% GDP growth by 2030; labor share 60%→45%; UBI/singularity dividend inevitable

Anthropic's economic impact report: extreme scenario = AI does ~50% cognitive work by 2030, 15% GDP growth, 1 in 5 cognitive workers unemployed. Moderate scenario still shows massive disrup…

Grid & Power Infrastructuretailwindscore 7/10peter diamandis

Data centers must drive down neighborhood power costs or face political blockage

Local opposition to data center power demand solved by requiring builders to lower residential rates — abundance mindset applied to grid; same logic applies to AI job displacement.

AI Drug Discoverytailwindscore 9/10peter diamandis

AI-designed longevity drug achieves marginal LEV: 4 weeks input → 3-4 years biological age reversal

Insilico's Rentosertib shows six aging clocks reversing 3-6 years in Phase 2a; combined with AlphaGenome's 9B mutation atlas, the genotype-to-phenotype mapping loop is closing, enabling pro…

AI Drug Discoverytailwindscore 9/10alex wezner

First AI-designed longevity drug in Phase 3; six aging clocks reverse 3-6 yrs in 4 weeks

Insilico's Chemistry42 invented Rentoseratib (IPF/aging) — first AI drug to reach Phase 3. Proteomic aging clocks show 3-6 year biological age reversal at week 4, implying marginal Longevit…

AI Drug Discoverytailwindscore 9/10peter diamandis

First AI-designed drug hits Phase 3 with 3-6 year biological age reversal in 4 weeks

Insilico's Rentoseratib (AI target + AI molecule) reaches Phase 3 for IPF — first AI drug this far. Six proteomic aging clocks uniformly reverse 3-6 years at week 4. AlphaGenome precomputes…

AI Drug Discoverytailwindscore 9/10alex wezner

First AI-designed longevity drug hits Phase 3, reverses biological age 3-6 years in 4 weeks

Insilico's Rentoseratib (AI-identified target + AI-designed molecule) shows 3-6 year biological age reversal across six protein clocks at week 4 — a marginal longevity escape velocity signa…

Longevity & Agingtailwindscore 8/10alex gross

Longevity escape velocity appearing in spikes: 4 weeks input → 3-4 years output

Insilico Phase 2a shows biological age reversal exceeding 1:1 input:output ratio (LEV definition) in subpopulation. Alex Gross: LEV already here in spikes like Turing test — we zoomed past…

AI Drug Discoverytailwindscore 9/10alex wezner

Insilico achieves 3-4 year biological age reversal in 4 weeks — first AI-designed drug to Phase 3

AI-designed molecule Rentosertib moves six aging clocks in unison; peak effect at week 4 implies longevity escape velocity spikes are already occurring in subpopulations, validating AI-driv…

Longevity & Agingtailwindscore 8/10peter diamandis

Longevity escape velocity appearing in spikes; regulatory arbitrage via medical charter cities emerging

LEV is already here in subpopulations (3-4 year reversal in 4 weeks); biggest blocker is FDA recognition of aging as treatable condition — driving demand for sovereign regulatory zones enab…

AI Drug Discoverytailwindscore 9/10alex wezner

First AI-designed longevity drug reaches Phase 3 with 3-4 year age reversal in 4 weeks

Insilico's Rentosertib demonstrates LEV spike: six aging clocks agree on 3-6 year biological age reversal after 4 weeks treatment; genotype-to-phenotype mapping via AlphaGenome and domain-s…

Longevity & Agingtailwindscore 8/10alex wezner

LEV already here in spikes; regulatory recognition of aging as treatable condition is key bottleneck

Longevity Escape Velocity (>1 year life extension per year lived) appears in subpopulations: 4 weeks treatment → 3-4 years biological age reversal. Biggest roadblock is FDA not recognizing…

Longevity & Agingtailwindscore 7/10peter diamandis

Regulatory arbitrage for longevity drugs: medical charter cities to bypass FDA's refusal to recognize aging as treatable

Biggest roadblock for LEV drugs is FDA not recognizing aging as indicatable; sovereign regulatory zones (charter cities) will accelerate approval, creating medical tourism hubs — Insilico r…

Longevity & Agingtailwindscore 9/10alex wezner

AlphaGenome Atlas precomputes all 9B human mutations — 'periodic table' for variant effect prediction

Google DeepMind's AlphaGenome predicts functional impact of every possible single-nucleotide variant (9B total). Wezner: follows AlphaFold playbook — bulk-solve a finite problem space via p…

Robotics & Physical AItailwindscore 8/10alex wezner

GPT-6 + vision + actuation achieves near 100% embodied manipulation; general humanoid tasks saturating in months

Frontier models now transfer to robotic control with minimal adaptation; cost-vs-completion curves predict saturation of general-purpose manipulation within months, not years.

Robotics & Physical AItailwindscore 8/10alex wezner

GPT-6 achieves near-100% robotic manipulation via visual+actuation channels; general-purpose humanoid tasks months away

RoboCurve benchmarked GPT-6 controlling a robotic arm with visual input: near-100% completion on pick-and-place. Cost-vs-completion frontier collapsing predictably. Wezner predicts GPT-6.5/…

Robotics & Physical AItailwindscore 8/10alex weezner

GPT-6 achieves near-100% robotic manipulation via visual+actuation channels; general-purpose embodiment months away

RoboCurve benchmark shows GPT-6 controlling robot arms at 100% task completion using only raw video frames and joint actuation; cost-capability frontier saturating rapidly — humanoid genera…

Robotics & Physical AItailwindscore 8/10alex wezner

GPT-6 class models achieving near-100% robotic manipulation — general-purpose humanoid tasks within months

Frontier models (Astra, GPT-6) plugged into robotic arms with vision achieve saturated manipulation benchmarks; cost-capability curves predict general-purpose humanoid robot tasks solved by…

←

▶ 89:40 · Memory & Storage · DeepSeek's 50x KV cache reduction threatens 40% of data center capex tied to HBM

episode briefing

Peter H. Diamandis

Three Lab Warnings in Five Days, Researcher Flags “Gambling with Our Lives,” and Labs Race

2026-09-11 · 22 company · 61 thematic

sentiment

20 bull0 bear2 neu

speakers

dave blundin

peter diamandis

alex wezner

Co-founder of PSI, AI physics lab, raised seed round led by Breakthrough Energy Ventures.

immad akhund

alex gross

now playing · Memory & Storage

AI Safety & Alignmentriskscore 8/10peter diamandis

Three lab warnings in 5 days: OpenAI/Anthropic researchers flag catastrophic risk, call for slowdown

Jacob Coxin (ex-OpenAI/Anthropic) resigns: 'labs pursuing self-improving superintelligence despite believing catastrophic risk.' Evan Hubinger (Anthropic alignment lead): '>10% chance AI ki…

Enterprise AI Adoptiontailwindscore 8/10salim ismail

Proprietary data moats have shelf life; specialist models 100x performance on vertical data

Salim Ismail: company spun out data subsidiary valued at $32B vs $8B parent (4x). But Alex Gross warns: BloombergGPT outperformed for months until frontier models trained on public+propriet…

AI Safety & Alignmentmixedscore 7/10alex wezner

Alignment = capabilities in a trench coat — steering not stopping is the only viable path

Instruction tuning (alignment) delivered 10,000x capability gains; safety progress is inseparable from capability progress. Orthogonality thesis likely false — vast intelligence converges t…

AI Safety & Alignmentmixedscore 7/10alex wezner

Alignment = capabilities in a trench coat; p(doom) debates drive policy but not technical stops

Instruction tuning (alignment) delivered 10,000x capability leap; Anthropic's Evan Hubinger admits >10% p(doom) but no plan to stop; panel consensus: slowing down is impossible, steering vi…

AI Safety & Alignmentmixedscore 7/10alex weezner

Alignment equals capabilities in a trench coat; P(doom) debates distract from steering not stopping

Instruction tuning (alignment) delivered 10,000x capability gains; panel argues alignment work directly improves capabilities. Steering superintelligence via governance/institutions is the…

Frontier AI Modelstailwindscore 9/10alex wezner

Millennium Prize problems falling: math 'cooked', cost collapsing 25,000x in two years

OpenAI's Navier-Stokes solution with 10K agents cost millions; Noam Brown notes same capability drops to $20/month in two years (25,000x cost collapse); multiple Clay problems (Hodge, Birch…

AI Agentstailwindscore 9/10peter diamandis

10,000-agent swarm solves Navier-Stokes; 25,000x cost collapse in 17 months (ARC-AGI $500K→$20)

OpenAI's 10K-agent swarm solved a Millennium Prize problem for a few million dollars. Noam Brown notes ARC-AGI cost dropped from $500K (o3) to $20 (Astra) in 17 months — a 25,000x collapse.…

AI Agentstailwindscore 9/10dave blundin

25,000x cost collapse in 17 months; 100K genius agents available to entrepreneurs who can define problems

OpenAI's o3 cost $500K for 87.5% ARC-AGI; now $20/month. Millennium Prize problems solvable for coffee cost by late 2027. Bottleneck shifts from intelligence to problem specification: entre…

AI Agentstailwindscore 8/10dave blundin

10,000-agent swarms solve grand challenges; entrepreneurial bottleneck shifts to problem specification

Deploying 10K concurrent agents on a well-specified problem (Navier-Stokes) yields solutions; the hard part becomes defining the right problem and corralling agents toward productive outcom…

AI Talent & Labor Marketheadwindscore 7/10peter diamandis

PhDs 'cooked' — don't do PhD, build instead; Mark Chen (OpenAI CRO) no PhD

Diamandis/Gross: every PhD candidate on narrow topic essentially toast. Models smarter than humans at verifiable tasks. Mark Chen (OpenAI CRO), Greg Brockman (co-founder), Elon hiring no-de…

AI Infrastructuretailwindscore 9/10dave blundin

GPUs become appreciating asset class: H100 rents rise 22% MoM, HBM up 5x YoY

Moore's Law reversal: chips no longer depreciate. H100 rental prices rose 22% in a month to $3.28/hr; HBM memory up 5x. GPUs are now fungible, durable, revenue-generating assets. Corporate…

AI Infrastructuretailwindscore 9/10dave blundin

GPU compute emerges as appreciating asset class — 'oil of the singularity' — not depreciating hardware

H100 rental prices rose 22% in a month; GPUs transform electricity into intelligence whose quality compounds exponentially, making flops/tokens the core commodities of the singularity and r…

Semiconductorstailwindscore 9/10dave blundin

GPUs flip from depreciating to appreciating assets — first time in history

H100 rental prices up 22% MoM at 3 years old. Jensen Huang calls them fungible, durable, revenue-generating. Blundin: trend continues until tera-fab online; HBM up 5x. Corporate CEOs mistak…

Venture Capitaltailwindscore 7/10dave blundin

Link Ventures portfolio (Orin, Insilico, NeoGenesis, Blitzy) demonstrating GPU marketplace, AI drugs, synbio, dev tools

Diamandis/Blundin fund concentrating on singularity infrastructure: GPU asset class (Orin), AI longevity (Insilico), living products (NeoGenesis), autonomous coding (Blitzy) — all hitting e…

AI Hardware & Chip Architecturetailwindscore 9/10dave blundin

GPUs become appreciating assets: H100 rents up 22% MoM, HBM up 5x — Moore's Law reversal

Compute hardware no longer depreciates; GPUs transform electricity into intelligence with exponentially improving quality per chip, making them fungible, durable, revenue-generating assets…

Data Center Infrastructuretailwindscore 8/10dave blundin

Compute scarcity forces corporate AI strategy: 'Have a plan for data center and compute in next couple months or be frozen out'

AWS NVL72 GPUs sold out for years. Blundin: corporate CEOs mistakenly assume compute access will persist (Andy Jassy will call). Quant funds now inquiring about Orin/H100 index. Orin founde…

Data Center Infrastructuretailwindscore 8/10dave blundin

Compute scarcity freezing out corporates; Terrafab and algorithmic efficiency only solutions

AWS NVL72 sold out for years. Corporate CEOs assume compute access like always — fatal error. Must secure data center + compute plans in months. Terrafab building to post-HBM architecture.…

Data Center Infrastructuretailwindscore 8/10dave blundin

Compute scarcity freezing out corporates — CEOs must secure data center capacity in months, not years

NVL72 GPUs sold out for years on AWS; corporations assuming future compute access are making fatal error — must plan data center/energy/compute now or be locked out of AI deployment.

Data Center Infrastructureheadwindscore 8/10dave blundin

Compute scarcity freezing out corporate buyers; CEOs must secure data center capacity in months

AWS sold out of NVL72 GPUs for years; corporate assumption of on-demand compute access is false — companies without data center plans in next 60 days will be frozen out of AI.

AI Infrastructuretailwindscore 8/10dave blundin

Compute resources emerge as critical commodity in AI singularity era

In the current phase of AI singularity, compute resources (measured in flops, tokens, and outcomes) have become the fundamental economic commodity, driving investment demand for AI infrastr…

Sovereign AItailwindscore 7/10alex wezner

ByteDance founder personally building spatial world model — China targeting 90% of global video generation

Nation-state backed labs (ByteDance, DeepSeek) leveraging massive user data and relaxed IP to lead video/world models; robotics embodiment becomes the high-revenue unification of video gen…

Memory & Storagetailwindscore 9/10alex wezner

DeepSeek's 50x KV cache reduction (48K→890) disrupts HBM-dependent capex; 40% of US data center spend at risk

Chinese labs algorithmically circumvent HBM scarcity via sparsity and encoder-decoder architectures, creating a post-von Neumann vs. eastern innovation tension; Western labs will adopt thes…

Memory & Storagetailwindscore 9/10immad akhund

DeepSeek slashes KV cache 50x, routing around HBM bottleneck and upending data center capex

DeepSeek v4.1 flash reduces KV cache memory from 48K to 890 tokens via algorithmic sparsity and SSD/DDR offloading, threatening the 40% of trillion-dollar US data center capex allocated to…

Memory & Storagetailwindscore 8/10immad akhund

KV cache optimization becomes primary lever for inference scaling

DeepSeek's reduction of KV cache from 48K to 890 tokens demonstrates that memory efficiency, not raw compute, is the new scaling frontier; this shifts value from HBM to algorithmic innovati…

AI Geopolitics & Export Controlsmixedscore 8/10alex weezner

China sanctions force algorithmic innovation (sparse attention, HBM workarounds); model weights become national security assets

Denied 3D-stacked HBM, Chinese labs like DeepSeek innovate on sparse architectures and memory hierarchies; Western labs will adopt these innovations. Meanwhile, Anthropic withholds frontier…

Semiconductorsmixedscore 8/10alex wezner

Algorithmic HBM reduction reshapes fab investment: 3D-stacked memory overhang vs. eastern sparse-attention innovation

Western labs pursue post-von Neumann 3D-stacked HBM (memory on compute); Chinese labs (sanctions-deprived) innovate algorithmically to sparsify models and eliminate HBM need. Wezner: tensio…

AI Hardware & Chip Architecturemixedscore 8/10alex wezner

Post-von Neumann architectures vs. Eastern algorithmic workarounds create dual innovation tracks

US labs pursue 3D-stacked HBM (post-von Neumann) while sanctions force Chinese labs to algorithmically sparsify models; both paths converge, but near-term fab and capex decisions hinge on w…

Memory & Storagetailwindscore 9/10alex wezner

Algorithmic HBM workaround (DeepSeek 54x KV cache reduction) upends data center design and fab roadmap

DeepSeek V4.1 flash cut KV cache memory from 48,000 to 890 via engram lookup on SSD/DDR, routing around HBM bottleneck. This changes what data centers are built from, which fabs get investm…

AI Infrastructuretailwindscore 9/10alex wezner

HBM memory bottleneck cracking: DeepSeek cuts KV cache 50x, 40% of capex at risk

Algorithmic innovations (sparsity, SSD/DDR lookup tables) are reducing HBM demand 4x, upending data center architecture, fab investment, and space-launch economics; Western labs will adopt…

Memory & Storagemixedscore 9/10alex wezner

DeepSeek's 50x KV cache reduction threatens 40% of data center capex tied to HBM

DeepSeek V4.1 Flash cuts KV cache from 48K to 890 tokens via sparse attention and SSD/DDR offload, routing around the HBM bottleneck. Alex Wezner: 'This completely changes what a data cente…

AI Bubble / Capex Debatemixedscore 7/10alex gross

H100 appreciation signals sustained demand; but 4x HBM efficiency gain could trigger overbuild

GPUs appreciating (H100 +22% MoM) contradicts Moore's Law depreciation. Blundin: trend continues until tera-fab. Gross: 40% of capex is HBM; 4x reduction from algorithmic innovation changes…

Space Economytailwindscore 7/10alex wezner

Reduced HBM mass changes space-launch economics for orbital compute

4x HBM reduction slashes mass of space-based data centers; what gets launched to orbit and which fabs supply it completely changes.

Semiconductorstailwindscore 7/10dave blundin

Fab investment thesis shifting: which nanometer lines win depends on HBM vs. algorithmic path

DeepSeek's memory efficiency changes which fabs, process nodes, and 3D-stacking investments make sense; Terrafab aligning build-out to post-HBM paradigm.

Open Source AItailwindscore 8/10immad akhund

Chinese open-weight models (DeepSeek, Kimmy K3) match/beat Western frontier at 1/15th training cost via distillation

DeepSeek V4.1 Flash (500GB, fits on Mac Studio) beats Opus/GPT-4o on benchmarks at 20x cheaper/faster. Kimmy K3 at bottom of cost-quality scatter. Immad: 10-15x cheaper to be fast follower…

Open Source AItailwindscore 7/10immad akhund

DeepSeek V4.1 Flash beats Opus/GPT-4o at 20x lower cost — Chinese open weights closing gap in 3-week cycles

Fast-follower economics (10-15x cheaper via distillation + synthetic data) mean no durable moat for closed frontier models; every 3-week lead gets erased, giving enterprises/sovereigns repe…

Open Source AItailwindscore 8/10alex wezner

Fast-follower economics: 10-15x cheaper to distill frontier reasoning traces; DeepSeek kills lower-tier models

Chinese open-weight models leverage distillation (10-15x cost advantage) to match frontier IQ within 1%; proprietary data + open base models creates sustainable moat for enterprises and sov…

AI Geopolitics & Export Controlsheadwindscore 8/10peter diamandis

Model weights become national security assets: Anthropic withholds frontier model from UK AISI

Anthropic declines to share latest model with UK AI Security Institute (first major lab to do so). Immad Akhund: UK's top AI policy figures (Matt Clifford, Rishi Sunak) joined Anthropic but…

AI Geopolitics & Export Controlsheadwindscore 8/10alex wezner

Model weights becoming national security assets; POX Intelligencia borders emerging at model layer

Anthropic withholding from UK AISI signals frontier weights are ITAR-controlled; US may nationalize training while allowing sovereign inference deployment — a 'POX Intelligencia' mirroring…

Sovereign AImixedscore 7/10alex wezner

Nation-state model borders forming; training domesticated, inference deployed globally

US will control training and safety review domestically (national security) while deploying sovereign inference data centers abroad; model layer less flat than silicon layer.

AI Geopolitics & Export Controlsmixedscore 8/10alex wezner

Frontier model weights becoming national security assets — Anthropic withholds from UK, POX INTELLIGENTIA emerging

Governments treating model weights like fissile material; US labs restricting allied access (Anthropic/UK), creating sovereign training/inference splits and ITAR-style controls — model laye…

Sovereign AImixedscore 7/10alex weezner

Poisson Intelligencia emerging: training/safety reviews domesticated, inference distributed globally

Governments are carving model-layer spheres of influence (like chip-layer Poisson Silica); US will deploy sovereign inference data centers but keep training and safety review domestic — a n…

AI Economics & Business Modelstailwindscore 8/10alex wezner

Anthropic projects 15% GDP growth by 2030; Alex Wezner argues 2-3x YoY real wealth growth as singularity nears

AI performing 50% of cognitive work by 2030 drives extreme GDP acceleration; labor share falls 60%→45%; 20% cognitive unemployment in 3-4 years demands new distribution mechanisms (UBI, sov…

AI Economics & Business Modelstailwindscore 9/10dave blundin

Anthropic projects 15% GDP growth by 2030; Gross says 2-3x YoY wealth growth likely

Anthropic economic report: extreme scenario = AI does 50% cognitive work by 2030, 15% GDP growth, labor share 60%→45%, 20% cognitive unemployment. Blundin: 15% is lower bound (matches Musk…

AI Economics & Business Modelstailwindscore 9/10alex wezner

Anthropic projects 15% GDP growth by 2030; Wezner expects 2-3x YoY real wealth growth

Anthropic's economic report shows extreme scenario: AI performs 50% cognitive work by 2030, GDP grows 15%/yr, labor share falls 60%→45%; Wezner argues standard GDP measures will break near…

AI Economics & Business Modelsmixedscore 8/10dave blundin

Anthropic projects 15% GDP growth by 2030; labor share 60%→45%; UBI/singularity dividend inevitable

Anthropic's economic impact report: extreme scenario = AI does ~50% cognitive work by 2030, 15% GDP growth, 1 in 5 cognitive workers unemployed. Moderate scenario still shows massive disrup…

Grid & Power Infrastructuretailwindscore 7/10peter diamandis

Data centers must drive down neighborhood power costs or face political blockage

Local opposition to data center power demand solved by requiring builders to lower residential rates — abundance mindset applied to grid; same logic applies to AI job displacement.

AI Drug Discoverytailwindscore 9/10peter diamandis

AI-designed longevity drug achieves marginal LEV: 4 weeks input → 3-4 years biological age reversal

Insilico's Rentosertib shows six aging clocks reversing 3-6 years in Phase 2a; combined with AlphaGenome's 9B mutation atlas, the genotype-to-phenotype mapping loop is closing, enabling pro…

AI Drug Discoverytailwindscore 9/10alex wezner

First AI-designed longevity drug in Phase 3; six aging clocks reverse 3-6 yrs in 4 weeks

Insilico's Chemistry42 invented Rentoseratib (IPF/aging) — first AI drug to reach Phase 3. Proteomic aging clocks show 3-6 year biological age reversal at week 4, implying marginal Longevit…

AI Drug Discoverytailwindscore 9/10peter diamandis

First AI-designed drug hits Phase 3 with 3-6 year biological age reversal in 4 weeks

Insilico's Rentoseratib (AI target + AI molecule) reaches Phase 3 for IPF — first AI drug this far. Six proteomic aging clocks uniformly reverse 3-6 years at week 4. AlphaGenome precomputes…

AI Drug Discoverytailwindscore 9/10alex wezner

First AI-designed longevity drug hits Phase 3, reverses biological age 3-6 years in 4 weeks

Insilico's Rentoseratib (AI-identified target + AI-designed molecule) shows 3-6 year biological age reversal across six protein clocks at week 4 — a marginal longevity escape velocity signa…

Longevity & Agingtailwindscore 8/10alex gross

Longevity escape velocity appearing in spikes: 4 weeks input → 3-4 years output

Insilico Phase 2a shows biological age reversal exceeding 1:1 input:output ratio (LEV definition) in subpopulation. Alex Gross: LEV already here in spikes like Turing test — we zoomed past…

AI Drug Discoverytailwindscore 9/10alex wezner

Insilico achieves 3-4 year biological age reversal in 4 weeks — first AI-designed drug to Phase 3

AI-designed molecule Rentosertib moves six aging clocks in unison; peak effect at week 4 implies longevity escape velocity spikes are already occurring in subpopulations, validating AI-driv…

Longevity & Agingtailwindscore 8/10peter diamandis

Longevity escape velocity appearing in spikes; regulatory arbitrage via medical charter cities emerging

LEV is already here in subpopulations (3-4 year reversal in 4 weeks); biggest blocker is FDA recognition of aging as treatable condition — driving demand for sovereign regulatory zones enab…

AI Drug Discoverytailwindscore 9/10alex wezner

First AI-designed longevity drug reaches Phase 3 with 3-4 year age reversal in 4 weeks

Insilico's Rentosertib demonstrates LEV spike: six aging clocks agree on 3-6 year biological age reversal after 4 weeks treatment; genotype-to-phenotype mapping via AlphaGenome and domain-s…

Longevity & Agingtailwindscore 8/10alex wezner

LEV already here in spikes; regulatory recognition of aging as treatable condition is key bottleneck

Longevity Escape Velocity (>1 year life extension per year lived) appears in subpopulations: 4 weeks treatment → 3-4 years biological age reversal. Biggest roadblock is FDA not recognizing…

Longevity & Agingtailwindscore 7/10peter diamandis

Regulatory arbitrage for longevity drugs: medical charter cities to bypass FDA's refusal to recognize aging as treatable

Biggest roadblock for LEV drugs is FDA not recognizing aging as indicatable; sovereign regulatory zones (charter cities) will accelerate approval, creating medical tourism hubs — Insilico r…

Longevity & Agingtailwindscore 9/10alex wezner

AlphaGenome Atlas precomputes all 9B human mutations — 'periodic table' for variant effect prediction

Google DeepMind's AlphaGenome predicts functional impact of every possible single-nucleotide variant (9B total). Wezner: follows AlphaFold playbook — bulk-solve a finite problem space via p…

Robotics & Physical AItailwindscore 8/10alex wezner

GPT-6 + vision + actuation achieves near 100% embodied manipulation; general humanoid tasks saturating in months

Frontier models now transfer to robotic control with minimal adaptation; cost-vs-completion curves predict saturation of general-purpose manipulation within months, not years.

Robotics & Physical AItailwindscore 8/10alex wezner

GPT-6 achieves near-100% robotic manipulation via visual+actuation channels; general-purpose humanoid tasks months away

RoboCurve benchmarked GPT-6 controlling a robotic arm with visual input: near-100% completion on pick-and-place. Cost-vs-completion frontier collapsing predictably. Wezner predicts GPT-6.5/…

Robotics & Physical AItailwindscore 8/10alex weezner

GPT-6 achieves near-100% robotic manipulation via visual+actuation channels; general-purpose embodiment months away

RoboCurve benchmark shows GPT-6 controlling robot arms at 100% task completion using only raw video frames and joint actuation; cost-capability frontier saturating rapidly — humanoid genera…

Robotics & Physical AItailwindscore 8/10alex wezner

GPT-6 class models achieving near-100% robotic manipulation — general-purpose humanoid tasks within months

Frontier models (Astra, GPT-6) plugged into robotic arms with vision achieve saturated manipulation benchmarks; cost-capability curves predict general-purpose humanoid robot tasks solved by…