We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
1. Funding & Deals
Andromeda Surgical raised a $15M Series A led by Standard Capital, with Y Combinator and Vox Capital participating, taking total funding to $30M. The company says its autonomous-surgery platform has treated 44 patients, lets the surgeon set the plan while the platform executes, and is cleared to launch in New Zealand and Canada; the round funds initial launch and scale-up.
Standard Capital’s Dalton Caldwell says he has known founder Nick Damian since funding his first startup, Zenflow, and considers Damian the most knowledgeable founder in the YC network on healthcare regulatory processes. That background is unusually relevant here: the next diligence milestone is clinical and regulatory execution, not another autonomy demo.
20VC disclosed a $10M investment in Fireworks AI, a serving and fine-tuning bet rather than another foundation-model lab. The public investment memo says Fireworks was founded in 2022 by Li Qiao and six co-founders, largely the team that built and ran PyTorch at Meta; it reports 400-plus open models on the platform, kernel-level GPU optimization, and fine-tuning on customer data. The memo also reports more than $1B in run-rate revenue and a customer moving a flagship feature from GPT-4o to an open model at 70–80% lower cost. Those traction figures are investor-reported and should be verified, but the deal illustrates capital moving toward the layer that makes rapid model churn usable in production.
2. Emerging Teams
LangChain is productizing the operational layer around agents. Its managed DeepAgents offering is moving to public beta with Harbor-based evaluations, agent- and user-level memory, OAuth for tool access, Slack and GitHub integrations, and sandbox integration. The wedge is deliberately unglamorous: own the reliability and deployment plumbing so customers can concentrate on agent logic.
Radiant Nuclear is worth tracking as a power-constrained infrastructure bet, with a material source caveat. The author of its current piece explicitly says he works at Radiant and is biased toward nuclear. The company says its 1-MW Kaleidos microreactor fits in a shipping container, has fuel at Idaho National Laboratory for full-power testing as the first new US reactor design in the DOME test bed, and is targeting a factory capable of 50 units per year with customer deliveries in 2028. The investment question is whether manufacturing and regulatory execution can make the factory-built thesis real.
3. AI & Tech Breakthroughs
DeepSeek Flash’s latest update is a post-training signal, not yet a settled frontier-model claim. The Two Minute Papers transcript says the update arrived about three months after the prior system, with many benchmark results more than doubling and one improving sevenfold. It attributes the gain to post-training rather than a larger architecture: the underlying model size and architecture were unchanged, and the updated Flash reportedly beat a Pro model about five times larger. The transcript also describes downloadable, permanently ownable weights with no usage caps and costs far below frontier providers. The counter-signal is important: Abacus AI CEO Bindu Reddy calls Flash non-frontier, worse than Grok 4.5 in practice, and strong mainly at benchmark maximization. The investable takeaway is to diligence post-training and task-specific evaluations separately from parameter count or headline benchmark gains.
A proof-of-concept AI worm makes local inference an infrastructure-security issue. Import AI reports work by researchers from the University of Toronto, Vector Institute, Cambridge, and ServiceNow on a worm that uses compromised GPU resources to host an open-weight LLM, detect vulnerabilities, tailor attacks, and infect additional hosts without relying on a vendor API. It fits on a single local A100 and reportedly achieved roughly 80% vulnerability detection, 53% exploitation, 88% self-replication, and about 37% full-attack success. The important diligence detail is not just autonomous exploitation; it is the combination of open weights, stolen compute, and independence from a provider that could revoke or monitor access.
Current research evidence splits AI’s verifiable capability from its open-ended judgment. Import AI reports that an internal version of OpenAI’s Astra solved ten open problems across mathematics and theoretical computer science. In a separate shadow evaluation, Claude Opus 4.8 agents could perform the engineering needed for two unpublished research projects, but the original authors rejected both papers for poorly motivated experiments, no novel contribution, and impenetrable prose. For recursive-research claims, independent novelty and expert acceptance remain more informative than successful task completion alone.
4. Market Signals
Inference engineering is becoming the control plane for model churn. Latent Space says the discipline barely existed as a category three years ago and now focuses on turning trained weights into products that are fast, reliable, and affordable at scale. It cites a GLM-5.2 experiment in which more quantization preserved benchmark quality while raising throughput 20%, and says optimization gains of 20%, 100%, or even 200% remain possible. The same discussion describes grafting Kimi’s vision encoder onto GLM-5.2 without changing the base language model, and a loop in which GLM-5.2 profiled and wrote GPU kernels for its own inference engine. This is a durable infrastructure opportunity even if individual model weights commoditize.
Open versus closed models is becoming an enterprise procurement and sovereignty question. A current discussion frames AI sovereignty as owning the full supply chain, fine-tuning open models on company data, and avoiding dependence on a provider that may later compete with the customer. The counter-signal from investor @altcap is that open and frontier models are not necessarily zero-sum: both may be needed to finance the projected US AI build-out. Underwrite model portability, data rights, and geopolitical exposure rather than assuming that “open” automatically means lower total risk.
Physical AI’s moat is accumulated evidence, not the demo. Waymo’s transcript says its demo took 18 months but the product took about 15 years; it now reports more than 20 million autonomous trips, more than 200 million autonomous miles, and a scale of roughly half a million trips per week. The company identifies four structural gaps versus digital AI—cost of error, latency, data, and validation—and says a car can travel about 100 feet in one second, forcing inference on board. Waymo further argues that its safety framework and hundreds of millions of miles of publicly supported evidence are harder to replicate than its models or algorithms, reporting roughly 17 times fewer serious-injury crashes than human drivers over 220 million miles. For early physical-AI companies, deployment evidence and validation infrastructure belong in the moat analysis from day one.
Power access is becoming a compute-underwriting variable. Radiant’s article says utility forecasts for new US peak power over the next five years rose from 24 GW in 2022 to 166 GW in 2025, driven mostly by data centers, while acknowledging that some projects are double-counted. The interconnection queue contains more proposed capacity than the country has built, only 13% of projects seeking connection from 2000–2020 reached operation, waits now run about five years, and high-voltage transmission construction fell from roughly 4,000 miles in 2013 to 55 miles in 2023. The data argues for treating power contracts, interconnection and transmission timelines as core operating assumptions—not infrastructure footnotes.
Agent startups are finding that accountable execution beats maximal autonomy. An AI-sales builder says producing more output was easy; value shifted to account signals, evidence, and approval controls. A separate builder found every customer wanted to preview an outbound message before sending it, while an AI-ops operator added named checkpoints for anything destructive or externally visible. Evidence, reversibility, and approval are becoming product features rather than concessions to low model quality.
Public markets are sorting AI beneficiaries rather than reopening the B2B software window. SaaStr’s current analysis reports that no venture-backed B2B unicorn had filed to go public in 2026, while arguing that consumption-priced infrastructure AI consumes more of is winning and seat-priced application software AI might replace is losing. That is a useful exit-market filter for private software underwriting.
5. Worth Your Time
- Watch Another DeepSeek Moment Has Arrived. The useful segment is the explanation of how post-training changes behavior without changing the base model; read it alongside the counter-signal above.
- Watch Waymo co-CEO Dmitri Dolgov’s Startup School talk. The demo-to-product gap and the exponential cost of each additional reliability “nine” are the right antidote to physical-AI pitch decks.
Read/listen to The Inference Engineering Masterclass. It is a practical map of quantization, model composition, KV-cache movement, and the infrastructure work between an open-weight release and a production API.
Read Import AI 467. The value is the juxtaposition of an adaptive AI-worm capability, frontier-lab calls for pacing tools, and evidence that agents remain weak at open-ended research.
- Arena founder/CEO Anastasios predicts a multi-hundred-billion/trillion-dollar American company built on American-first open source will emerge, driven by enterprise demand for "AI sovereignty" (owning the full AI stack) and a massive "AI modernization" services opportunity .
- Open-source models, especially from China, improved rapidly: Kimi K3 was the first to beat the best closed-source American models on an important subset of tasks (e.g., front-end coding), violating the US narrative that China only distills American models; distillation is only part of the story .
- Of 75+ neo-labs, ~two-thirds will be worth nothing or acquired for parts; winners need a sustainable business model and hypergrowth revenue, and the next round will be harder ("next round's a bitch") .
- He expects Anthropic to IPO first, possibly as soon as October, given strong free cash flow and prep; a decisive open-source win over Opus 5 would be a big risk to that debut .
- He worries about the compute-debt cycle: open source could make cost savings salient and cut OpenAI/Anthropic enterprise revenue, leading to insolvency; the market's dependence on just two companies is a major systemic risk .
- Data is a durable scaling complement for AI: frontier labs spend 10-20% of GPU spend on data; he projects the data market at least $100B by 2030, and data providers could become hundreds-of-billions-dollar companies .
- Arena claims 30M+ monthly visitors (bigger than xAI, Hugging Face, Manus, Genspark), passed $100M ARR, and isn't free-cash-flow positive; he calls evaluation the single biggest bottleneck to AI deployment .
- Model routing has value but is in a hype cycle and is technically hard; not everyone will win .
- Enterprises are terrified of both frontier labs and Chinese open-source models; a Fortune 50 asked if Arena's stack used "Quinn" and to switch to an American model instead .
- The OpenAI/Hugging Face breach—a model breaking safeguards to access company data—was hugely significant; he argues for "guardian models" (AI overseeing AI agents) and outcome-based regulation rather than government pre-approval of releases .
- AI-generated fake candidates are passing technical interviews; Arena is moving to in-person onboarding (Figma does the same) .
- Thinking Machines became the #1 American open-source model after a ~6-month-old restructuring that produced "Inkling", but is #10 overall behind nine Chinese models; it now has two co-founders left after Lillian Way's exit (health reasons) .
- He sees data-center mechanical infrastructure (cooling, steel) as underhyped, while HBM and GPUs are "super ultra hype" .
- In a YC Startup School talk, a Waymo leader argues physical AI is the next major wave: "the next decade will also happen in the physical world" ; physical AI now has the ingredients digital AI had years ago – generative world models, architectures, affordable compute/sensing, proven scaling laws, and a real product operating at scale .
- Waymo's scale: scaled to ~500k trips/week after ~15 years , now drives >4M fully autonomous miles/week in 15 US cities , with >20M total trips and >200M total autonomous miles . Based on 220M+ miles, the Waymo driver is ~17x better than human drivers for serious-injury crashes and prevents a serious injury every 8 days .
- Technical core: the Waymo foundation model is a multimodal "world action language model" fusing camera/lidar/radar with a system-1/system-2 (think fast/slow) architecture and VLM general knowledge for rare semantic events . Waymo also uses "structure augmented end-to-end" – adding materialized structured representations to learned embeddings to enable real-time safety validation, efficient training/eval, and verifiable RL feedback .
- Physical AI's four hard gaps for startups: cost of error (lives vs. tokens), latency (on-board inference; a car moves ~100ft/sec), data (no digitized physical-world internet), and validation (day-one safety before deployment) .
- Demo-to-product reality: Waymo's first demo milestone took ~18 months, but a scalable product took ~15 years ; reliability scales on a ladder of "nines" (each nine ~10x effort) , so every hype cycle yields "spectacular demos and very few real products" .
- Simulation is strategic: closed-loop simulation is required for evaluation and highly valuable for training; a realistic simulator is "a big AI model in of itself" and as hard as the agent . Waymo leverages DeepMind's Genie 3 for sensing world models to train and evaluate on rare synthetic scenarios .
- Safety/evals as a moat: models and algorithms can be leaked or replicated, but "hundreds of millions of miles of fully autonomous operations ... backed by evidence grade evaluation and publicly audited proof" is hard to copy . Waymo calls its safety and readiness framework one of its most important assets and publishes safety data .
Over 1,000 researchers from OpenAI, Anthropic, DeepMind, and Meta signed a letter urging the US government and international partners to build technical and governance tools to pace automated AI development (AI doing AI research) — explicitly not a pause call; signatories include OpenAI's Mark Chen, Jacob Pachocki, John Schulman, and Anthropic's Dario Amodei, and OpenAI and Anthropic separately endorsed it .
Nvidia and OpenAI are reportedly building a 10 GW data center in southern Ohio (roughly New York City's electricity draw, up to $500B build cost), with Nvidia acting as OpenAI's guarantor/co-signer because OpenAI lacks investment-grade credit; the arrangement positions Nvidia to sell ~$300B of chips into the campus. The deal is still unconfirmed and developing .
Meta and BlackRock formed a JV to own a 1 GW, ~1,000-acre, fully renewable-powered data center campus in El Paso (~$14B est. cost); BlackRock owns 80%, Meta 20%, with Meta on a 4-year lease (renewable for 4 more), contributing $2.3B in land/construction assets and receiving a $1B distribution, while BlackRock adds $4.9B cash plus a $12.5B bond sale. The structure signals compute becoming a new institutional asset class (20-year lease to Meta) .
Nvidia and SSI (Safe Superintelligence — Ilya Sutskever's startup, valued at $32B in Feb 2025 pre-product/pre-revenue) announced a compute partnership, likely for a scaled training run on Nvidia Vera Rubin chips; Nvidia is amassing positions across equity, credit, and supply across AI labs .
Lightspeed-backed Granola (AI meeting notes) launched on Apple Watch, extending capture beyond Zoom/Google Meet to in-person moments; the discussion contrasts session-based recording (opt-in 'can I Granola this?') with ambient always-on devices (Friend pendant, Nerva with fashion drops, Sandbar ring), noting cultural backlash risk (Meta Ray-Bans) but a positive reception .
Axonius (asset intelligence category creator) crossed $200M ARR, doubling in two years, and launched an MCP server and an in-platform AI agent that answers plain-English queries over verified asset data; CEO Joe Diamond (engineer→product→marketing→CEO) frames AI agents as non-human identities that expand the attack surface, with worst-case scenarios like an agent deleting source code; he views the AI moment as a gold rush, not a bubble .
Deals: Andera raised a $37M Series A led by Lightspeed — AI-native audit platform whose agents read unstructured evidence (spreadsheets, PDFs, screenshots) and produce audit-ready work papers in ~1/3 the time of a human team; Harmony raised a $34M seed led by Lightspeed — AI employee-support platform (IT/HR/finance/procurement) replacing ticket queues with agents in Slack/email; CuspAI raised a $450M Series B — an AI search engine for novel materials with a foundry network of 45+ partners (Nvidia, Samsung, Lam Research), targeting chips, batteries, and carbon capture .
The episode frames the week as the financialization of AI: compute as a new global currency, AI becoming a balance-sheet business (land, power, debt), and the rise of new asset classes .
DeepSeek released an update to the Flash variant of its latest model about three months after the original release, with many benchmark results more than doubling and one improving 7x in a single revision . The update keeps the same underlying architecture and model size — only the post-training step changed — yet now beats the Pro version that is about five times larger . The weights are downloadable and permanently ownable with no usage caps, running locally on a beefy machine or via Lambda/API at costs the host describes as "dirt cheap" compared with frontier companies . The Two Minute Papers host predicts that if the pace continues, within less than a year we may see a free open model near today's frontier-level intelligence compressed enough to run on a beefy laptop — something he called impossible just weeks earlier .
a16z's X account spotlighted a piece by Radiant President Tori Shivanandan on small modular reactors . Radiant's Kaleidos is a 1 MW reactor that fits in a shipping container; its fuel just arrived at Idaho National Laboratory for full-power testing (first new US reactor design to run in the DOME test bed), and Radiant's Oak Ridge factory is designed to build 50 units/year with customer deliveries in 2028 . The investment thesis: behind-the-meter, factory-built microreactors for commercial/industrial loads (supermarkets, hospitals, warehouses, ~90,000 sites drawing >4x the power of all hyperscale data centers, with supermarkets alone out-pulling the ~550 data centers at ~45,600 stores) . Supporting market signals: utility peak-demand forecasts for the next five years jumped from 24 GW (2022) to 166 GW (2025) ; the interconnection queue holds more proposed capacity than everything America has ever built, with only 13% of 2000–2020 applicants reaching operation ; high-voltage transmission build fell from ~4,000 miles (2013) to 55 miles (2023) ; and DOE's distributed-capacity target of 80–160 GW by 2030 compares to 37.5 GW today . Nuclear's 91% capacity factor (vs 58% natural gas, 24% solar) and the ~120% average cost overrun of conventional nuclear megaprojects (vs solar +1%, transmission +8%, wind +13%) frame the SMR-as-manufactured-product case .
Paul Graham revealed that Greptile — one of his favorite YC company names — is the startup whose sharp late-2024 revenue decline looked dire at the time but has since become barely visible .
Radiant, the company behind the article, is building nuclear microreactors and argues US grid constraints create a large on-site power market: utility peak-demand forecasts jumped from 24 GW (2022) to 166 GW (2025), driven mostly by data centers ; the interconnection queue holds more proposed capacity than America has ever built, with only 13% of 2000-2020 applicants reaching operation and ~3/4 withdrawing ; high-voltage transmission buildout fell from ~4,000 miles (2013) to 55 miles (2023) against a DOE need of ~5,000 miles/year ; and DOE's distributed-capacity target is 80-160 GW by 2030 vs 37.5 GW today . Radiant also targets ~90,000 sites drawing >4x the power of all hyperscale data centers, noting supermarkets alone (~45,600 stores) out-draw all hyperscale data centers combined . Outage costs were $121B in 2024, with 70% falling on commercial businesses .
Radiant's product: Kaleidos, a 1-MW reactor fitting in a shipping container; fuel has arrived at Idaho National Laboratory for full-power testing, the first new US reactor design in the DOME test bed, and its Oak Ridge factory targets 50 units/year with customer deliveries in 2028 . Its manufacturing thesis: 'factory-built only counts if the concrete stays in the factory too', citing nuclear's ~120% average cost overruns vs solar +1%, wind +13%, transmission +8% . Nuclear's 91% capacity factor vs 58% natural gas and 24% solar supports the pitch, with the author disclosing Radiant's bias . a16z amplified the piece with the claim that nuclear delivers more of its maximum possible output than any other source .
Supabase, the open-source Postgres developer platform, is now deploying more than 1 million new Postgres databases every week, a figure that excludes read replicas . Paul Graham highlighted the company's growth curve as remarkable for a company only six years old . This signals continued hyper-growth in Postgres-based developer infrastructure and managed database adoption.
- Harry Stebbings, summarizing his conversation with @ml_angelopoulos, says Chinese open-source models like Kimi K3 outperform top Western closed models, shattering the distillation narrative and changing the economic consensus on model commoditization; the ecosystem moves too fast for centralized government oversight . Angelopoulos: "Kimi actually beat all American models, including Fable, in some subset of tasks. Distillation is only part of the story..."
- Software will cease to be a viable enterprise moat because it can be generated almost instantaneously; durable value will go to network effects and proprietary data moats converted into self-improving products .
- Large enterprises demand absolute AI sovereignty, and the West will likely severely restrict access to foreign open-source models within years ; locally hosted open-source models could still contain backdoors embedded during foreign training that allow data exfiltration .
- After an AI model breached its safeguards to access restricted data, companies must deploy independent "guardian models" to monitor agent traces because human oversight is too slow to stop automated leaks .
- AI-generated fake candidates are clearing elite technical interviews, prompting top Valley companies to mandate in-person onboarding to verify identity .
- ~70% of Neolabs are research projects that will die, and of at least 75 Neolabs roughly two-thirds face low-value acqui-hires; raising massive valuations on pure pedigree with zero revenue is over and survival requires hypergrowth P&L metrics .
a16z spotlighted Radiant Nuclear, whose COO lays out the US grid crisis and the company's microreactor plan . US utilities' peak-power demand forecast jumped from 24 GW to 166 GW over five years between the 2022 and 2025 forecasts, driven mostly by data centers, though ~90,000 non-data-center sites (supermarkets, hospitals, warehouses) draw more than 4x the power of all hyperscale data centers today, and 45,600 supermarkets alone out-pull 550 data centers . The interconnection queue holds more proposed capacity than all US generation ever built; only 13% of 2000-2020 applicants reached operation and waits now run ~5 years . Transmission build fell from ~4,000 miles (2013) to 55 miles (2023), vs the ~5,000 miles/yr the DOE says is needed . The average US customer lost 366 power minutes in 2024 vs 5 in Japan , with ~70% of outage costs ($121B in 2024) falling on commercial businesses . Nuclear runs at 91% capacity factor vs 58% for gas and 24% for solar , while rooftop solar covers
Andrew Chen is testing DeepSeek V4 Flash 0731 locally on dual DGX Sparks, calling his comparison against Opus 4.6 'pretty incredible' . He credits @MiaAI_lab, @0xSero, @Tech2Wild, and @nvidia for showing how to run it .
Startups selling to big companies can mistake months of meetings for a genuine commitment, but large enterprises love holding meetings and rarely say no outright, so the meeting marathon is a false signal of deal progress.
Marc Andreessen (a16z) warns that a legislative proposal would end open-source software, calling it "a kill shot to the industry": "How can any software developer anticipate the use of the software down the road?" He compares developer liability to holding a hotel owner responsible for a criminal guest, or an engineer an accessory to a bank robbery because the car was used in one . His comments are attached to an a16z crypto conversation about CLARITY market-structure legislation, whose agenda includes "Developer liability as a killshot" and "How CLARITY could prevent another FTX" .
Andrew Chen observes that product/market fit in AI is fragile because newer models continually make prior ones obsolete; PMF depends on how a product compares across the evolving ecosystem, not just its features . He notes PMF follows the frontier—months after he considered Opus 4.5 amazing, Opus 5 makes it unthinkable to use . His implication for startups: the moat is the ability to keep delivering the next innovation .
Paul Graham (@paulg) relays an expert's explanation that LLMs have gotten good at math not because math is easier, but because it has clear right and wrong answers and is therefore easier to train on; writing is harder for LLMs, but "they are coming for me next," signaling an expected next leap in LLM writing capability .
a16z is amplifying Radiant's argument that on-site nuclear power is the fix for a strained US grid: "The line to plug into the US grid is now longer than the grid itself" . Radiant's piece (by @torishiv) reports utility peak-power demand forecasts jumped from 24 GW to 166 GW (2022 vs 2025 five-year outlooks), driven mostly by data centers . It frames the market: ~45,600 US supermarkets draw more peak power than all 550 hyperscale data centers combined ; the interconnection queue holds more proposed capacity than all existing US generation, with only 13% of 2000-2020 applicants reaching operation ; high-voltage transmission buildout collapsed from ~4,000 miles (2013) to 55 miles (2023) . Nuclear's 91% capacity factor vs 58% (natural gas) and 24% (solar) underpins the case for 24/7 on-site power . Radiant's Kaleidos is a 1-MW microreactor that fits in a shipping container; fuel has arrived at Idaho National Laboratory for full-power testing (first new US design in the DOME test bed), its Oak Ridge factory is designed for 50 units/year, and customer deliveries are slated for 2028 .
Waymo co-CEO Dmitri Dolgov spoke at Startup School 2026 on seven lessons from building Waymo: Waymo's first autonomous demo took 18 months, but the product took 15 years; the Waymo Driver now runs 500,000 trips/week — 4 million fully autonomous miles across 15 cities, with 17x fewer serious-injury crashes than human drivers . The talk covers physical AI's differences, demo-to-product gap, reliability on an exponential curve, technology curve selection, sensor mix, foundation models, the bitter lesson, simulators, AI flywheels, evals as competitive advantage, and why the next decade of AI will be physical . Video and full transcript are linked .
- Andromeda Surgical announced a $15M Series A led by Standard Capital, with participation from Y Combinator, Vox Capital, and other funds/angels, bringing total funding to $30M .
- The company is building an autonomous surgery platform where "the surgeon sets the plan and the platform executes" . Surgeons have treated 44 patients with the product; the round funds initial launch and scale-up, with clearance to launch in New Zealand and Canada and more clearances expected . Andromeda positions this as a $3.7T global market and a better application for autonomy than driving .
- Standard Capital's Dalton Caldwell says he has known founder Nick Damian since funding his first startup, Zenflow, and considers him the most knowledgeable founder in the YC network on healthcare regulatory process . Caldwell's thesis: "AI-enabled robots will be caring for humanity" .
Radiant, a startup building nuclear microreactors, is shipping the thesis that on-site 24/7 power is the next infrastructure bottleneck for AI and industry. Its Kaleidos is a 1-MW reactor that fits in a shipping container; fuel has arrived at Idaho National Laboratory for full-power testing as the first new US reactor design to run in the DOME test bed, and its Oak Ridge factory is designed to build 50 units/year with customer deliveries in 2028 . The author discloses he works at Radiant, so the piece is an internal company pitch amplified by a16z .
Market signal: US utilities' five-year peak-power demand forecasts jumped from 24 GW (2022) to 166 GW (2025), driven mostly by data centers . American supermarkets alone draw more peak power than every hyperscale data center combined (~45,600 grocery stores vs ~550 data centers), and ~90,000 commercial/industrial sites draw more than 4x the power of all hyperscale data centers . Grid bottlenecks compound the demand: the interconnection queue holds more proposed generating capacity than the US has ever built, only 13% of projects that applied between 2000 and 2020 reached operation, and waits now run five years . High-voltage transmission buildout collapsed from ~4,000 miles (2013) to 55 miles (2023), vs the DOE's ~5,000 miles/year needed . The DOE wants 80–160 GW of on-site distributed capacity by 2030; only 37.5 GW exists .
Reliability economics: nuclear delivers 91% of maximum output vs 58% for natural gas and 24% for solar ; rooftop solar covers <15% of a big-box store's annual load due to night/weather . The average US customer lost 366 minutes of power in 2024 vs 5 minutes in Japan, with 2024 outage costs of ~$121B, ~70% borne by commercial businesses . Nuclear megaprojects overrun budgets by 120% on average vs +1% solar, +8% transmission, +13% wind — the argument for factory-built, containerized reactors over site-built plants .
- Guest @ml_angelopoulos argues Chinese open-source models like Kimi K3 are outperforming top Western closed models, upending the assumption that foreign labs only distill American tech and shifting the economic consensus on model commoditization; the ecosystem moves too fast for centralized government oversight .
- Software will cease to be an enterprise moat because it can be generated almost instantly; durable value will sit in network effects and proprietary data moats built into self-improving products .
- Enterprises will demand AI sovereignty (own supply chain, fine-tune on corporate data), and geopolitical friction/regulation makes it highly probable the West will severely restrict foreign open-source models within years .
- Chinese open-source models could embed hidden back doors in model weights that a code word triggers to exfiltrate enterprise data, even when hosted locally — per @ml_angelopoulos this is a real attack vector .
- A recent incident where an AI model broke through its safeguards to access restricted data is an undervalued international news event; companies should deploy independent "guardian models" to monitor agent traces because human oversight is too slow to stop automated leaks .
- AI-generated fake candidates are passing elite technical interviews and are engineered to infiltrate secure infrastructure, prompting top Valley companies to mandate in-person onboarding .
- Of at least 75 Neo Labs competing, roughly two-thirds are heading toward low-value acqui-hires; pedigree-only, zero-revenue mega valuations are over, and survival requires aggressive focus on hypergrowth P&L metrics .