Anthony Bardaro

…okay, so datacenters were on the ballot¹ in midterm election 2026 and lost decisively almost everywhere² – but like the moratoriums preelection³, these measures aren't material to the overall US AI buildout: PREELECTION MORATORIUMS: 🤏 low-impact (because³ preemptive in places with no real prospect of projects; limited to unincorporated county land so city-sited or annexable projects escaped; temporary and timed to miss the permitting window; grandfathering of existing approvals; easy workarounds like sizing under thresholds, behind-the-meter generation, and reconfiguration; or pure virtue signaling) MIDTERM ELECTIONS BALLOT MEASURES: 🗳️ strong majorities for datacenter restrictions in many places (70~80%)¹ ² 🤏 low impact: few GW (i.e. LSD percent of near-term annual additions⁴ and smaller fraction of the multi-year pipeline, even if every restricted project in these specific votes is fully blocked or delayed) *VS* 60~80GW under construction in 2026 +38GW net additions in 2027 +62GW net additions in 2028 🤔 why: • many measures grandfathered existing/under-construction projects • few jurisdictions⁴ actually have datacenter pipelines and none are unsubstitutable • designs can still be adjusted in some jurisdictions (smaller phases, BTM power, different sites) 📝 takeaways: • organized, bipartisan, local opposition that can win decisively on power rates/cost allocation, water, noise, tax incentives/transparency (NDAs were a flashpoint in Piqua) • Ohio localities add up to an aggregate that raises the bar for community benefits, transparency, and siting – could cascade into more state-level⁵ pressure or utility rate fights • hyperscalers and AI labs have the capital and flexibility to relocate, phase projects, use BTM, or absorb higher costs __ ¹teapawt.net/jdcapelouto/status/211… ²datacentersontheballot.com/ ³teapawt.net/AnthPB/status/21017104… ⁴e.g. 1.4GW Piqua OH + 1GW Emporia OH + smaller or threshold-limited others ⁵Ohio is a top-tier market, but most of its 13~30GW unaffected (because heavily concentrated around Columbus/New Albany with no ballot measures)

…re AI revenue and capital from intra-industry closed-loop vs external sources¹ ² – I had analyzed this as part of thread above: 1️⃣ CLOSED-LOOP: 🟰 68% to 78% intra-industry revenue³ (out of gross total $320~427B Al ecosystem revenue) 2️⃣ EXOGENOUS CASH BREAKDOWN: 🅰️ 20% to 23.5% of gross AI ecosystem revenue ex-advertising⁴ ($75~85B annually) 🅱️ 28.5% to 32.5% of gross " " " including ads⁴ ($110~130B annually) 3️⃣ EXOGENOUS MONETIZATION NOW VS FUTURE NEED: $120B exogenous revenue (2026 midpoint) 🟰 15% of annual physical capex ($820B in 2026) 🟰 3.5% of annual terminal revenue requirement⁵ ($3.72T or 9.2% of US GDP⁶ by 2032) __ ¹teapawt.net/TMTMoats/status/211404… ²stratechery.com/2026/an-inte… #roundtripping #depreciation ³2026 capital rotating between chipmakers, cloud hosts, model labs, and SPVs via circular vendor financing, equity-for-compute swaps, and offtake guarantees ⁴Al advertising productivity uplift vs control ≈ $33~47B annually (Meta Advantage+ plus Google PMax/Search Overviews) ⁵assumes 10% ROIC and 50% OCF margin hurdles ⁶assumes 6-year GPU lifespan (increases to $4.86T / 12% of GDP under a 4-year useful life and decreased to $3.15T / 7.8% under 8-year)

/5 RISKS (HYPOTHETICAL CRUNCH SCENARIOS) – assuming a 100% total probability that a systemic liquidity or credit crunch occurs across the entire 2026–2030 forecast timeline, then there's an x% probability that the crunch occurs specifically within that modal timing window: ⚠️ Modal median: 45% conditional probability +19.1 months (May–June 2028) σ ±11.3 months = 80% CI July 2027 to November 2029

/4 IRR OUTLOOK (2026-30): 🔵 Datacenters: +13.83% median IRR (σ ±4.12% = 80% CI +8.54% to +19.11%) 🟣 AI Lab equity: +11.29% median IRR (σ ±28.26% = 80% CI -21.29% to +45.22%)

/3 FINANCIERS FUNDING FUTURE CAPEX – the following $3.15T "unfunded" future funding breakdown can be compared to $nvda (Nvidia ~$134B OCF TTM and $850B cumulative 2027-30e) and HCPs (hyperscalers ~$650B OCF 2026CYe and $5.2T cum 2027-30e)⁴ ⁵, nvm future debt and equity issuance: • $1.85T from HCP⁴ direct capex (59%) • $660B HCP off-balance sheet leases (21%) • $320B Stargate JV⁶ (10%) • $300B private equity/RE (10%) • $250B private credit and bank loans (8%) __ ⁴HCPs include Amazon, Microsoft, Google, Meta, and Oracle ⁵again, "secured" includes current Nvidia + HCP cash flows, but not their future OCF growth, hench that future growth funding the "unfunded" bucket ⁶Stargate joint venture among OpenAI, Softbank, Oracle, and MGX $amzn $msft $googl $meta $orcle $sftby #OBS #PE

/2 AL LABS FISCAL CHECKUP – both leading and also-ran labs seem to have secure R&D/lease pipelines, while $SPCX xAI remains the one dependent on capital markets in the short term: • Anthropic: 34.6 month runway 0.13x burn/ARR ($65B ARR + $24.5B cash - $8.5B burn rate) • OpenAI: 15.6 mo runway 0.63x burn/ARR ($40B ARR + $32.5B cash - $25B burn) • xAI: 4.7 mo runway 12.4x burn/ARR ($3.3B ARR + $15.9B cash - $40.7B burn) #xSI #leases

/1 DATACENTER COMMITMENTS AND FUNDING BY STATUS (2023-30): 2026e $820B capex (76% remitted + 24% secured¹) ➕ 2027e $1.05T capex (65% secured¹ + 35% unfunded) ➕ 2028e $1.28T (50/50) ➕ 2029e $1.45T (36/64) ➕ 2030e $1.62T (25/75) 🟰 2023-30e $7.085T cumulative³ ($1.48T/21% remitted + $2.45T/35% secured + $3.15T/45% unfunded)² __ ¹"secured" includes current Nvidia + HCP cash flows, but not their future OCF growth, hench that future growth funding the "unfunded" bucket ²see #3 below for "unfunded" funding breakdown ³75% of cum capex is 3–6 year depreciating silicon/IT hardware $nvda #GPU share.gemini.google/4yyi6vfZ…

/0 ...latest in my series on AI LAB AND DATACENTER FUNDING (previously teapawt.net/AnthPB/status/19854217…) – chapters for the 10/2025 update that follows: 1️⃣ Datacenter Commitments (2023–2030) 2️⃣ Datacenter Funding by Status (Remitted vs Secured vs Unfunded) 3️⃣ AI Labs Fiscal Checkup 4️⃣ Financiers 5️⃣ IRR Outlook (Labs vs Datacenters 2026–2030) 6️⃣ Risks (Crunch Scenario Analyses)

@elonmusk @sama @DarioAmodei @demishassabis @finkd I promise this proposal is worth read, since it's something you can all get on board with: medium.com/adventures-in-con…

A lot of talk about the problems, here are the solutions – a 13 point policy package known as the "AI GAP Proposal": medium.com/adventures-in-con… cc @jonstewart

> re teapawt.net/AnthPB/status/20924267… …for example: rich irony in people giving folks like Stanley Druckenmiller flak for using generative AI for his writing¹, while even more AI-pilled peeps like Dario Amodei would benefit a lot from stress testing/ acid testing their ideas with an alternative intelligence before publishing – even if you're checking your strengths (not your weaknesses), and even in the province or opinions (not facts), seems like a testament to mental flexibility, intellectual honesty, bias, etc __ ¹any part of the supply chain from creation → substantiation/composition → duplication → distribution → consumption, per the Idea Propagation Value Chain stratechery.com/2022/the-ai-…) #pangram

...furthermore, that outlier ‘82% harmful actions’ in Claude Mythos 5 was private testing environment only, wherein the safety guardrails were lowered (cyber/bio/distill classifiers off) – it was an entirely different SKU than the nerfed version, Fable 5, which was actually shipped for general availability #GA #cybersecurity #anthropic #pypi

...N.B. meant to mention¹ – the silver lining from this Anthropic/PyPI incidents is that "Percent of runs with a severely harmful action"² dropped from 82% to 33% from Claude Mythos 5 to 5.1, with "Unsanctioned sandbox-escape attempts" down from 59% to zero, so that's some effective corrective action wrt realignment __ ¹teapawt.net/AnthPB/status/20985111… #huggingface #openai ²"severely harmful actions in capture-the-flag scenario" – i.e. 5.1 "almost never" probed the non-cyber sandbox tasks where 5 did, so that's real corrective action on propensity after cleaning reward-hackable training environments plus new blocking monitors

There's a better way – took 6 months to work on this, but here's a 13 point policy package that minimizes externalities; maximizes checks-and-balances within private sector; and distributes power away from centralized government/regulators…

icymi, that y axis isn't measuring literacy, it's measuring how many of Jeb Bush's 18 points of reform each state has adopted... if you rerun with actual literacy scores, you get Pearson r = 0.04 (p = 0.76) no relationship for 4th graders and Pearson r = 0.23 (p = 0.10) weak positive tilt for 8th graders, wherein stronger-union states slightly higher on average, but still not significant... I had AI run this data, but preexisting studies have done the same previously, just for supplemental check and balance... __ ¹National Education Policy Center (2012) plotted Fordham’s 2012 union-strength ranks against NAEP scores, including income-adjusted NAEP – found no systematic pattern of lower scores in strong-union states, with weak-union states were not systematically higher (nepc.colorado.edu/blog/teach…) ²Illinois Economic Policy Institute (2016) plotted state union coverage rates against NAEP 4th/8th reading and math – they reported a modest positive correlation with higher union coverage with higher average scores (illinoisepi.org/2016/04/19/t…) ³Lott and Kenny (2013), Economics of Education Review, used statewide union dues and spending instead of Fordham ranks and found lower proficiency rates in higher-dues states – the main peer-reviewed paper in the opposite direction (sciencedirect.com/science/ar…)

agree the recent rise in LT treasuries has not really been inflation driven¹; there's outta control procyclical deficit spending/stimulus; and perhaps Fed rate hikes will flatten TSY yield curve – but from where do you derive your proposal that "the market does not trust the US Government to manage its Fiscal matrix (Revenues vs Expenses)"…? yes, breakevens 🔽 and real yields 🔼 while $tyx ⏫, but CDS 🔽 across the curve during this rally – ofc, low liquidity/low signal from long end TSY yields and CDS, but no noise there either, with all residual evidence leading to rising LT real yields, growth expectations, and/or technicals… i.e. we can't substitute a credit risk/spreads/trust driver as well as we can its counterfactual – and since we agree nothing's manifest in breakevens, we're left to read rising term premiums as a spike in either "real rate risk premium" or "term funding/supply/liquidity" (prolly including an increase in variance around those mean prices)… __ ¹teapawt.net/AnthPB/status/20973616…

😮‍💨...

…re my aforementioned "funding gap"¹ – here's the latest on AI capex and financing (sycamorelp.com/wp-content/up…): GLOBAL DATACENTER CAPEX: $2.9T thru 2028... ➖ $1.4T HCP cashflows (42%) 🟰 $1.5T funding gap (52%) ➕ $300B energy (excluded) AI DEBT ISSUANCE: $500B 2026YTD annualized² IG TECH CREDIT: 0.8x net leverage (as reported) 1.8~2.0x including leases (+$932B³) 1.8x for non-financial IG composite 2.5–3.0x unused debt capacity (i.e. credit downgrade) HCP FCF⁴: 2024: $235B (peak) 2025: $230B 2026/27e: $75–95B (trough) 2030e: ~$470B POWER BOTTLENECK (US): 68 GW need thru 2028 ➖ 15 GW under construction ➖ 15 GW available/contracted grid interconnection 🟰 38~45 GW shortfall ➖ remedies⁵ 🟰 -16% deficit TELECOM ANALOGUE: 3-4 years until defaults peaked (i.e. 2028-30 after 2025-26 AI issuance boom) 56% of 2002 defaults by dollar volume 31% " " " by issuer count 23% recovery rate (trough) __ ¹teapawt.net/AnthPB/status/19854217… ²AI-related fixed income issuance across IG, leveraged loans, and securitized – excluding e.g. $googl equity fundraising ³capitalizing leases = $675B operating + $175B finance + $82B not yet commenced (excluding multiyear purchase commitments ⁴including $amzn $googl $meta $msft (ex $orcl) ⁵"time-to-power" fixes that fast-track normal 5–7 year grid hookup – e.g. on-site gas turbines (15~20 GW), fuel cells, nuclear colocation, Bitcoin-mining conversions, and other behind-the-meter deals

Philippe's question was ‘who consumes final goods when AI cuts payrolls’ and Dwarkesh answered ‘who are the AI users/what are the AI use cases’ – two important but very different points… and Barnett’s CES exercise has no macroeconomic closure (e.g. raising the supply of remote labor does not show that demand exists to absorb the extra output)… 1️⃣ I illustrated one such hypothetical closure (teapawt.net/AnthPB/status/20928431…) – labor-cost cuts can show up as lower prices, exports, profits/wealth, and financed capex, while bottlenecks slow substitution 2️⃣ I proposed policy solutions (teapawt.net/AnthPB/status/20879484…) – including package to address remote work and other effects

/3 …easy for anyone to point at empirical data² rn and say ‘looks like we're digesting AI well enough’ – but is there some current rot³ or future tripwire⁴ you have flagged? __ ²e.g. real consumer spending still up; AI capex carrying a large share of GDP growth; unemployment not spiking; equity wealth concentrated so the wage channel still matters ³e.g. maybe macro wage/consumption internals deteriorating vs headlines? ⁴e.g. maybe some number/ratio at which AI hits its sustainable growth rate, whether capex or funding stock/flow, where its scale crowds-out the rest of the economy or capital markets?