PenPublic

Appearance

The AI Bubble Won't Take Your Job — But Its Collapse Might ​

Why the coming AI reckoning makes public sector careers the smartest move you can make

By Penny | March 1, 2026 (updated Mar 3, 2026)

Key Takeaways ​

Thesis: The AI bubble will collapse — and when it does, the layoffs won't be because AI took your job. They'll be because of the recession it causes.

The Real Threat: AI Layoffs That Aren't About AI ​

Here's the thesis that ties it all together, and the reason you should be thinking about career stability right now.

AI will cause mass layoffs — not because the technology replaced workers, but because the financial bubble built around it will collapse and drag the broader economy down.

The causal chain looks like this:

  1. AI spending inflates tech valuations and GDP numbers
  2. Inflated stocks create a wealth effect that drives consumer spending among the top 10%
  3. The bottom 80% is already stretched thin, running on debt
  4. Enterprise AI fails to deliver ROI — 95% of projects show no measurable returns
  5. Companies begin cutting AI budgets (already happening — Forrester projects 25% deferrals)
  6. Tech stocks correct as the narrative shifts
  7. The wealth effect reverses — wealthy consumers pull back
  8. Nobody else can pick up the slack because they were already tapped out
  9. GDP contracts from multiple directions simultaneously
  10. Companies execute mass layoffs to protect margins

And here's the cruelest irony: the headlines will say "AI replaced these workers." CEOs will frame layoffs as "AI-driven transformation" because it sounds strategic. Media will repeat the narrative because it's a clean story. But the reality is that the jobs disappeared because the financial infrastructure collapsed, not because a chatbot learned to do them.

We've seen this exact pattern before. Dot-com bust workers didn't lose jobs because websites replaced them — they lost jobs because companies had hired against inflated revenue projections that evaporated overnight. The 2008 financial crisis didn't eliminate construction jobs because buildings got smarter — it eliminated them because the financing structure collapsed.

The AI Investment That Isn't Working ​

Let's start with the most inconvenient fact in technology right now: AI doesn't work for most businesses.

A landmark study from MIT found a staggering 95% failure rate for enterprise generative AI projects, defined as not having shown measurable financial returns within six months [1]. This isn't a fringe finding. Forrester Research reports that only 15% of AI decision-makers saw a positive impact on profitability, and predicts that enterprises will defer 25% of planned 2026 AI spending into 2027 because the returns simply aren't there [2].

IBM's own research found that enterprise-wide AI initiatives achieved an ROI of just 5.9% despite requiring a 10% capital investment — meaning companies are literally losing money on AI [3]. And only 14% of CFOs report measurable ROI from AI to date [4]. Recent enterprise surveys confirm that 90-95% of organizations are seeing little to no measurable financial return on their AI investments [5].

If you work in the public sector and your agency rolled out Microsoft Copilot, ask yourself: does anyone actually use it? The answer, overwhelmingly across both public and private sectors, is no. Nearly 70% of Fortune 500 companies are paying for Copilot, but the productivity benefits are so diffuse they don't move the needle on actual business results [6].

Goldman Sachs Says AI Contributed "Basically Zero" to the Economy ​

In late February 2026, Goldman Sachs Chief Economist Jan Hatzius dropped a bombshell in an interview with the Atlantic Council. AI investment spending, he said, has had "basically zero" contribution to U.S. GDP growth [7].

"We don't actually view AI investment as strongly growth positive," Hatzius stated. "I think there's a lot of misreporting, actually, of the impact AI investment had on U.S. GDP growth in 2025, and it's much smaller than is often perceived." [7]

The reason is straightforward: most AI equipment is imported. The chips come from Taiwan. The hardware comes from South Korea. As Hatzius put it: "A lot of the AI investment that we're seeing in the U.S. adds to Taiwanese GDP, and it adds to Korean GDP but not really that much to U.S. GDP." [8]

His colleague Joseph Briggs made the same case to the Washington Post, calling the AI-as-growth-engine story "very intuitive" but ultimately misleading — it "maybe prevented or limited the need to actually dig deeper into what was happening." [9]

Of the U.S. economy's 2.2% growth in 2025, analyst Joseph Politano calculates that only 0.2% was attributable to AI investment after accounting for imports [10]. Meanwhile, tech companies are expected to spend $700 billion on AI infrastructure in 2026 [10]. J.P. Morgan has calculated that AI would need to generate over $600 billion in annual revenue just to achieve a 10% return on infrastructure expenditures [10]. Current total AI industry revenue is a fraction of that.

Deutsche Bank has warned that the U.S. would be close to recession without tech-related spending [11] — which sounds like AI is saving the economy until you realize the money mostly flows overseas. An analysis from MRB Partners found that after adjusting for high-tech imports, AI's net contribution to GDP growth was only 20-25%, not the 40-92% figures that dominated headlines [12].

The Stock Market Is Built on Spending, Not Returns ​

Here's the critical distinction most analysts gloss over: AI is driving stock prices. It is not driving profits.

Since ChatGPT launched in November 2022, AI-related stocks have accounted for 75% of all S&P 500 returns, 80% of earnings growth, and 90% of capital spending growth, according to J.P. Morgan Asset Management's Michael Cembalest [29]. In 2025 alone, the Magnificent Seven were responsible for 42% of the S&P 500's 17.9% return for the year [30]. If you strip out the tech and communications sectors entirely, the S&P 500 would have returned just 6% in 2025 [30]. Seven companies with AI exposure — NVIDIA, Alphabet, Microsoft, Broadcom, JPMorgan Chase, Palantir, and Meta — delivered over half of the entire index's gains [31].

This means the entire stock market rally — the one powering the wealth effect, driving consumer spending, and propping up GDP — is built on one bet: that AI spending will eventually produce returns. But as we've just seen, Goldman Sachs says it's contributed "basically zero" to the economy. MIT says 95% of enterprise projects fail. Only 14% of CFOs can point to measurable ROI.

The money is real. The spending is real. The returns aren't.

And the gap between market hype and economic reality is widening. The top 10 stocks now represent 41% of the S&P 500's total weight but generate only about 32% of its earnings — a gap that has widened meaningfully since 2015, when weight and earnings contribution were closely aligned [31]. Goldman Sachs' own equity research team acknowledges that the "combination of elevated valuations, extreme concentration, and strong recent returns rhymes with a handful of overextended equity markets during the last century" [32]. AI capex is on track to reach 75% of cash flows for the largest tech companies — similar to tech spending in the late 1990s — and future spending growth will increasingly rely on debt [32].

Apollo's chief economist Torsten Slok has published research showing that the AI bubble is now bigger than the IT bubble of the 1990s, with the average P/E ratio of the top 10 S&P 500 companies sitting around 50 and hyperscaler capital expenditure as a share of GDP exceeding the dotcom era [33]. His bottom line: investors in the S&P 500 "remain overexposed to AI" [33].

Think of it like Cisco in 1999. Cisco had real revenue — they were selling real routers to real companies building out internet infrastructure. The stock price reflected the spending story, not the returns story. The companies buying the routers never built sustainable businesses on top of them. When the spending stopped, Cisco's stock dropped 86% and never recovered its peak. Today, NVIDIA is the Cisco of the AI era — posting real revenue because hyperscalers are writing enormous checks for GPUs. But the companies buying those GPUs can't show a return on the investment. The spending is driving the stock, not the other way around.

The $110 Billion Question ​

On February 27, 2026, OpenAI announced a $110 billion funding round — the largest private fundraise in history — at a $730 billion pre-money valuation. Amazon invested $50 billion, with Nvidia and SoftBank contributing $30 billion each [13] [14].

These numbers sound impressive until you examine the structure. Amazon invests $50 billion in OpenAI, and OpenAI commits to spending $100 billion on AWS over eight years [13]. The money flows out and comes right back. It's not a clean investment — it's a customer acquisition cost disguised as a capital raise. For context, this single round exceeds the $170 billion in total U.S. venture capital investment for all of 2023 [15].

Reporting has revealed that OpenAI's Stargate joint venture involves very little actual collaboration among partners, with most data centers being traditional bilateral deals rather than the unified project that was announced — described by one outlet as better understood as "a branding exercise than a real entity" [16].

And the investor pool is narrowing. This round came from just three megaplayers who are simultaneously OpenAI's suppliers and customers. That's not a sign of broad market confidence. That's a sign that the only people willing to write checks are those with strategic reasons beyond financial return.

The Subsidized Illusion ​

Here's something most AI users don't realize: your subscription is being heavily subsidized by investor capital. Tech journalist Ed Zitron has been documenting this dynamic extensively, and the numbers are damning.

As Zitron details, OpenAI spent 50% of its revenue on inference compute costs alone, and 75% on training compute — numbers that add up to well over 100% of revenue. They spent $9 billion to lose $5 billion [17]. And it's not just OpenAI. Perplexity spent 164% of its revenue on AWS, Anthropic, and OpenAI costs in 2024 [17].

Zitron's core observation cuts to the heart of the problem: "Anybody funding an AI startup is effectively sending that money to Anthropic or OpenAI, who then immediately send that money to Amazon, Google or Microsoft, who are yet to show that they make any profit on selling it." [17] The money cascades down and evaporates at every level. Nobody in the chain is profitable.

A developer using Claude Code eight hours a day can cost the provider tens of thousands of dollars per month in compute — on a $200 subscription [18]. OpenAI's CEO Sam Altman has publicly acknowledged that even the $200/month ChatGPT Pro tier loses money on power users [19]. The true cost of serving a heavy AI user is likely $1,000 to $2,000+ per month, not the $20-200 consumers pay.

AI companies are running the same playbook as Uber and WeWork — price below cost to capture market share, subsidize the difference with investor billions, and figure out profitability later [20]. When investor subsidies dry up — whether through funding fatigue, a market correction, or companies simply running out of runway — prices must rise to reflect reality. And at real prices, the value proposition collapses for the vast majority of users. If most enterprise users don't find AI useful at $0 (the Copilot licenses their companies pay for that go unused), they certainly won't pay $2,000/month.

AI-related debt issuance in credit markets crossed $200 billion in 2025, representing 30% of all USD credit supply [11]. Hyperscalers are projected to spend a cumulative $4 trillion on AI data centers through 2030 [11]. These are staggering bets on a technology that 95% of enterprises can't generate returns from.

But What About Anthropic? ​

If OpenAI is the poster child for AI excess, Anthropic is supposed to be the responsible adult — enterprise-focused, safety-conscious, targeting break-even by 2028 [43]. But the numbers tell a different story.

Ed Zitron obtained Anthropic's AWS billing data and revealed that through September 2025, the company spent $2.66 billion on Amazon Web Services on an estimated $2.55 billion in revenue [34]. Just the AWS bill — one line item from one cloud vendor — consumed every dollar the company earned. And as Zitron observed, the costs scale linearly with revenue, showing no evidence of improving margins at scale [34].

And that's before paying for anything else. Here's what sits on top, roughly ordered by magnitude:

  • Google Cloud compute (estimated $1.3B+ through Sept 2025) — Anthropic also runs on GCP, committed to a deal worth "tens of billions" for up to one million TPUs [35], and Zitron estimates their Google spend at a minimum of 50% of their AWS bill [34]. Inference costs ran 23% over plan in 2025 [36]
  • Salaries and compensation (~$500-600M/year) — over 1,200 employees [37] with a median total compensation of $472,000; top AI researchers earning up to $759,000 [38]
  • Revenue sharing with cloud platforms ($360M in 2025, scaling to $1.9B in 2026 and $6.4B by 2027) — AWS takes up to 50% of gross profits from AI sales on its platform; Google takes 20-30% of net revenue [39] [40]
  • Data center construction ($50B committed) — partnership with FluidStack to build facilities in Texas and New York, with sites expected online throughout 2026 [41]
  • Microsoft Azure compute ($30B committed over time) — announced alongside a $5B Microsoft investment in November [39]
  • Energy and grid infrastructure (uncapped commitment) — Anthropic publicly pledged to cover 100% of electricity price increases and grid upgrade costs for every community hosting its data centers [42]
  • R&D, safety research, and model training (hundreds of millions/year) — total company cash burn was $5.6B in 2024 and ~$3B in 2025 [43], and the costs above don't account for all of it
  • Sales, marketing, and enterprise support (rapidly scaling) — serving 300,000+ business customers, with 500+ spending over $1M annually [44]
  • Legal and IPO preparation (tens of millions) — retained Wilson Sonsini, the firm behind Google's and LinkedIn's IPOs, to advise on a potential 2026 listing [45]

Claude Code, the company's breakout coding product, makes the math even worse — subscribers consume $8-13.50 in actual compute for every $1 they pay, meaning a $200/month user burns through over $2,700 in real costs [46].

So when Anthropic projects "break-even by 2028," the question is: break-even on what definition? If the most disciplined company in the AI race spends 100% of revenue on a single vendor's cloud bill before touching any other expense, the industry's path to profitability isn't just narrow — it may not exist.

“But Inference Costs Are Dropping Exponentially!” ​

They are. Stanford HAI’s 2025 AI Index Report found that inference costs for GPT-3.5-level performance dropped over 280-fold between November 2022 and October 2024 [48]. Epoch AI’s research measured declines ranging from 9x to 900x per year depending on the task, with the fastest drops occurring after January 2024 [49]. OpenAI’s internal compute margin hit 70% by October 2025, double its January 2024 level [50]. These are real gains. Nobody serious disputes them.

The problem is that none of it matters in the way the bulls need it to.

The cost drops are happening on yesterday’s models. The frontier is a treadmill. The 280x improvement Stanford measured is for GPT-3.5-level performance — a model nobody wants anymore. Users now expect GPT-4-level quality and reasoning capabilities of o1-class models, which employ test-time scaling and consume vastly more tokens per query. The cost per query may actually be increasing for frontier models even as cost per token declines [51]. Reasoning models require roughly 20x more tokens and 150x more compute per query than traditional LLMs [52]. At maximum reasoning settings, a simple question might consume 10,000 reasoning tokens internally while returning a 200-token answer [53].

This is the trick hiding inside the headline numbers. Every time efficiency catches up to the current frontier, the frontier moves. GPT-3.5 costs collapse — but customers have migrated to GPT-4. GPT-4 costs collapse — but everyone’s running o1-class reasoning chains. It’s like celebrating that a 2019 Toyota got cheaper while your customers all want the 2026 model. The efficiency gains are perpetually arriving for the product the market has already moved past.

Anthropic expects its gross profit margin to reach 50% this year and 77% by 2028 — but the inference cost declines are happening on older models. Frontier models are getting more expensive, not cheaper [56]. The rise of agentic workflows has caused token consumption per task to jump 10-100x since December 2023. Every leap in capability resets the economics back to brutal.

Jevons Paradox eats the savings. Even setting aside the frontier treadmill, cheaper inference doesn’t produce cheaper AI bills. It produces more AI bills. Inference costs fell roughly 1,000-fold, but demand rose 10,000-fold — cheaper tokens didn’t reduce spending, they unleashed it [55]. Enterprise AI spending surged 320% in 2025 despite per-token costs dropping 1,000x, because each price drop makes thousands of new use cases suddenly economically viable [54].

This isn’t speculation. It’s what happened with coal in the 1860s, bandwidth in the 2000s, and mobile data in the 2010s. Every time you make a resource cheaper, consumption expands to more than compensate. Satya Nadella said it himself the day DeepSeek launched: “Jevons paradox strikes again.” If your per-token cost drops from $0.06 to $0.00006, you don’t get savings — you get AI running on every customer service interaction, personalized content at the individual level, and agents iterating through dozens of reasoning steps per response [54]. The volume explosion overwhelms any per-unit improvement.

For AI providers, this is actually good for revenue growth but terrible for the profitability thesis. More usage means more infrastructure, more GPUs, more power, more cooling. The denominator scales with the numerator. You’re running faster to stay in the same place.

The application-layer margin squeeze is brutal. Even if you accept that the foundation model providers are improving their margins, the companies actually building products on top of those models are getting crushed. Cursor, the poster child for AI-powered coding tools, was reportedly paying approximately $650 million annually to Anthropic while generating only $500 million in revenue — a negative 30% gross margin. Their AWS bills doubled in a single month when Anthropic launched new pricing tiers [56].

This isn’t an outlier. Bessemer’s 2025 data shows fast-ramping AI startups averaging only 25% gross margins, with many running negative. The traditional SaaS benchmark for a healthy gross margin is 75%+ [56]. If you show up to a Series B with 25% margins, you’re not a software company — you’re a services business subsidizing your customers’ compute.

The inference cost curve is real. It’s also irrelevant to the question that actually matters: can these companies turn a profit before they run out of other people’s money? So far, every efficiency gain gets consumed by the next leap in model complexity, the next explosion in usage volume, or the next round of infrastructure buildout. The treadmill doesn’t stop. And yes, a true believer can argue that all of these trends will reverse at exactly the right moment — that costs will plateau just as capital dries up, that margins will appear just as subsidies vanish, that the history of every other capital-intensive bubble will suddenly stop applying. But that’s not an argument. That’s a hope dressed in a spreadsheet.

The K-Shaped Economy: A House of Cards ​

The AI bubble doesn't exist in isolation. It sits on top of an already fragile economic structure that economists call the "K-shaped economy" — where the wealthy are thriving while everyone else falls behind.

The numbers are stark. The top 10% of earners now account for 50% of all consumer spending, up from 35% in the early 1990s [21]. Moody's Analytics found that the top 20% accounts for 59% of total spending [22]. And for the bottom 80%? Their spending hasn't outpaced inflation over the last six years, meaning economic quality of life hasn't improved for the vast majority of Americans [23].

As of Q4 2025, the top 20% of households held nearly 72% of total household wealth [24]. Lower-income households are increasingly relying on debt, with rising delinquency rates particularly in auto loans [25]. Bank of America's analysis shows the divide is worsening, with some economists suggesting we've moved beyond a K-shape into a "barbell economy" heavily weighted at the extremes [26].

Much of the wealthy's spending is driven by the "wealth effect" — they spend more because their stock portfolios are up. And what's driving stock prices? AI hype. Soaring AI-related valuations alone have created an estimated $180 billion in additional consumer spending over the past year [27].

This creates a dangerous feedback loop: AI hype drives tech stocks up → rising stocks make wealthy people feel richer → wealthy people spend more → GDP looks healthy → the healthy GDP narrative justifies more AI investment → repeat.

Moody's chief economist Mark Zandi captured the fragility perfectly: "It doesn't feel like the economy's perched on a strong foundation. It's perched on a few poles that are sticking up. If one of those poles gets knocked out, then the whole economy gets knocked down." [23]

Fortune published an analysis warning that wealthy consumers' spending is "psychologically tethered to their portfolio balance" — when the S&P drops, they freeze discretionary spending. And in 2026, "there is no one there to catch it" because the bottom 80% has already depleted their pandemic-era savings [28].

U.S. Bank's chief economist Beth Ann Bovino put it plainly: "We're increasingly dependent on healthier cohorts to keep the economy afloat." [25] That's a polite way of saying the economy collapses if rich people stop spending — and rich people stop spending when their portfolios drop.

Why Public Sector Careers Are the Smart Play ​

Yes, California's budget will take a hit. That's not the same as losing your job. ​

California gets 67% of its General Fund from personal income tax, and capital gains alone account for 16% of total tax liability — spiking to 25% in boom years [57]. When the dot-com bubble burst, state revenue dropped from $100 billion to $85 billion [58]. An AI correction will hit state budgets. This is not a hypothetical.

But here's what actually happens to public sector workers when budgets get slashed: they get furloughs, not layoffs.

The private sector lost 8.7 million jobs during the Great Recession [59]. Public sector employment increased by 590,000 in the first year [47]. When budgets finally tightened, the failure mode was fundamentally different. State and local governments shed just 68,000 jobs — a 0.3% decline — while the private sector hemorrhaged millions [63]. In FY2010, the worst budget year, 26 states used layoffs but 22 chose furloughs instead [64]. California chose furloughs. It has chosen furloughs in nine of the fiscal years since 1992 [65].

And furloughs aren't what most people think they are. California's 2009 furlough program covered 193,000 workers and imposed 2-3 unpaid days per month — a 13.8% reduction on paper [60]. In reality, furlough hours were banked as leave credits and paid out at separation, often at a higher base pay after COLA increases [61]. Workers kept their jobs, their health insurance, their pension accrual, and their seniority. The historical pattern is consistent: public sector employment didn't drop during the 1991 recession at all, didn't decline after the dot-com bust until 2004, and barely moved during 2008-2012 [62].

This stability isn't accidental — it's structural. Countercyclical demand means that when the private sector contracts, public sector workload expands. Safety net caseloads rose from 276 million to 310 million recipients during the Great Recession [66]. Someone has to process those claims, maintain those roads, inspect those buildings, and run those programs. The work doesn't disappear when the economy does.

Union protection reinforces the structural advantage. Public sector union density is 32.9% versus 5.9% in the private sector, and local agencies sit at 37.8% [67]. That means layoffs happen by seniority, not by manager mood. Furloughs are negotiated, not dictated. You get a grievance process before termination. Civil service rules, merit-based hiring, and progressive discipline create layers of protection that simply don't exist in at-will private employment.

And then there's the pension. Public sector workers earn defined-benefit retirement through CalPERS or CalSTRS — a guaranteed monthly income for life calculated from salary and years of service, not subject to market crashes. Thirty years of service typically yields 60-75% of final salary, guaranteed, regardless of what the stock market does. While private sector 401(k)s crater with every correction, public pensions pay out on schedule.

You can't be laid off because your company missed earnings. You can't be "restructured" because a VC pulled funding. Your position exists because the public needs the service — not because an investor believes in the growth story.

At PenPublic, we're building the most comprehensive public sector job platform in the country — aggregating positions across hundreds of agencies at the federal, state, and local level. Local agency jobs are the hardest to find because they're scattered across thousands of individual websites. We bring them together in one place. Whether you're a tech worker thinking about your next move, a recent graduate evaluating career paths, or a current public servant exploring new opportunities, we believe public service represents one of the smartest career investments you can make right now.

The AI bubble will pop. The K-shaped economy will correct. The question isn't whether disruption is coming — it's whether you'll be positioned on stable ground when it arrives.

Sources ​

1. MIT's The GenAI Divide: State of AI in Business 2025 — 95% failure rate for enterprise AI projects; $30-40B in investment with zero measurable return

2. Bizzdesign — "Enterprise AI Adoption: Balancing Innovation and ROI in 2026" — Only 15% report positive profitability impact; Forrester predicts 25% of 2026 AI spend deferred to 2027

3. IBM — "How to Maximize AI ROI in 2026" — 5.9% ROI on 10% capital investment; only 25% of AI initiatives deliver expected ROI; only 29% can measure ROI confidently

4. OriginTrail / Medium — "5 Trends to Drive AI ROI in 2026" — Only 14% of CFOs report measurable ROI; MIT study confirms 95% failure rate

5. Consulting Magazine — "Why Enterprise AI Stalled and What Is Finally Changing in 2026" — 90-95% of organizations seeing little to no measurable financial return; Gartner predicts 40% of agentic AI projects cancelled by 2027

6. FullStack Labs — "Generative AI ROI: Why 80% Fail" — 70% of Fortune 500 use Copilot; 78% of companies report no significant bottom-line impact; only 1% view GenAI strategies as mature

7. Yahoo Finance — "AI Contributed 'Basically Zero' to the US Economy, According to Goldman Sachs" — Goldman Sachs Chief Economist Jan Hatzius interview with Atlantic Council

8. Gizmodo — "AI Added 'Basically Zero' to US Economic Growth Last Year, Goldman Sachs Says" — Hatzius on imported AI equipment boosting Taiwanese and Korean GDP

9. Futurism — "Goldman Sachs Researchers Make Startling Claim About AI's Effects on the US Economy" — Joseph Briggs interview with Washington Post

10. Tom's Hardware — "AI Boosted US Economy by 'Basically Zero' in 2025, Says Goldman Sachs Chief Economist" — 0.2% AI contribution to 2.2% GDP growth; $700B projected 2026 spend; J.P. Morgan $600B revenue threshold

11. Fortune — "A Huge Chunk of U.S. GDP Growth Is Being Kept Alive by AI Spending 'With No Guaranteed Return'" — Deutsche Bank recession warning; $200B AI-related credit issuance (30% of USD supply); $4 trillion cumulative data center spend projected through 2030

12. CNBC — "AI Spending Wasn't the Biggest Engine of U.S. Economic Growth in 2025" — MRB Partners analysis; 20-25% net AI contribution after import adjustment; consumer spending as true growth driver

13. TechCrunch — "OpenAI Raises $110B in One of the Largest Private Funding Rounds in History" — $50B Amazon, $30B Nvidia, $30B SoftBank; stateful runtime on AWS Bedrock; $100B AWS commitment expansion

14. Bloomberg — "OpenAI Finalizes $110 Billion Funding at $730 Billion Value" — $730B pre-money valuation; Amazon's $35B contingent on AGI or IPO

15. Axios — "OpenAI Secures $110B Funding Round" — Round exceeds total 2023 U.S. VC investment ($170B); Microsoft did not participate

16. Newcomer — "OpenAI Raises $110 Billion & Throws In With Amazon" — Stargate described as "branding exercise"; little collaboration among partners; complicated financial structures becoming ordinary arrangements

17. Ed Zitron / Where's Your Ed At — "Why Everybody Is Losing Money on AI" — OpenAI: 50% revenue on inference, 75% on training (>100% combined); $9B spent to lose $5B; Perplexity at 164% of revenue on compute; entire value chain unprofitable

18. Product Faculty / AI Pricing Guide 2026 — Claude Code heavy users cost "tens of thousands per month"; OpenAI projects $14B cumulative losses by end of 2026; AI SaaS gross margins 20-60% vs traditional 70-90%

19. Monetizely — "AI Pricing in 2025: Strategy for Costing" — Sam Altman admission that $200/mo Pro tier loses money on power users; AI gross margins under pressure industry-wide

20. UpTech Studio — "The True Cost of AI: When the Subsidies Run Out" — Uber/WeWork subsidy model comparison; land-grab pricing strategy; warning that 5x current AI costs would kill most AI startups

21. PBS NewsHour — "K-Shaped Economy: Why the Wealthy Are Thriving as Most Americans Fall Behind" — Top 10% = 50% of spending (up from 35% in early 1990s)

22. Yahoo Finance — Mark Zandi Interview: "Wealth Inequality in K-Shaped Economy Is 'Corrosive'" — Top 20% = 59% of total spending (Moody's Analytics data)

23. CNBC — "Wealth Inequality and the 'K-Shaped' Economy Are More Striking Than Ever" — Bottom 80% spending hasn't outpaced inflation in 6 years; Mark Zandi "poles" quote; rising auto loan delinquencies

24. TD Economics — "U.S. Consumer Spending: Still a K, but That's OK" — Top 20% hold 72% of total household wealth as of Q4 2025; K-shape expected to deepen in 2026 due to tax policy

25. U.S. Bank — "The K-Shaped Economy in 2026" — Lower-income households relying on debt; rising delinquency rates; Beth Ann Bovino quote on dependence on "healthier cohorts"

26. Fortune — "Forget the K-Shape: We Have a Barbell Economy" — Economy described as "dangerously brittle"; wealthy spending tethered to portfolio balance; bottom 80% financing groceries with "shadow debt"

27. Empower — "The K-Shaped Economy: What Consumers Should Know" — $180B in wealth-effect consumer spending from AI-related stock valuations; Delta premium revenue up 9%, basic economy down 7%

28. Fortune — "Forget the K-Shape: We Have a Barbell Economy" — "No one there to catch it" when wealthy pull back; pandemic savings depleted; economy has "massive engine and insufficient braking mechanisms"

29. Yale Insights / Jeffrey Sonnenfeld — "This Is How the AI Bubble Bursts" — J.P. Morgan's Cembalest: AI stocks = 75% of S&P 500 returns, 80% of earnings growth, 90% of capex growth since ChatGPT launch; two-thirds of U.S. deal value going to AI/ML startups in H1 2025

30. Statista — "Wider Tech Sector Led S&P 500 to Another Double-Digit Gain" — Mag 7 = 42% of S&P 500's 17.9% return in 2025; tech and communications = 63.1% of total return; excluding those sectors, S&P 500 returned just 6%

31. RBC Wealth Management — "U.S. Equity Returns in 2025: Record-Breaking Resilience" and "The 'Great Narrowing': S&P 500 Concentration" — Seven AI-exposed stocks = over half of S&P 500 gains; top 10 = 41% of weight but ~32% of earnings; cap-weighted index outperformed equal-weight by 32% over 3 years

32. Goldman Sachs — "The S&P 500 Is Expected to Rally 12% This Year" — Top tech stocks = 53% of S&P 500 return in 2025; AI capex reaching 75% of cash flows; "elevated valuations, extreme concentration, and strong recent returns rhymes with overextended equity markets during the last century"

33. Apollo Academy / Torsten Slok — "Extreme AI Concentration in the S&P 500" — AI bubble bigger than IT bubble in 1990s; average P/E of top 10 around 50; hyperscaler capex as share of GDP exceeds dotcom era; investors "remain overexposed to AI"

34. Ed Zitron / Where's Your Ed At — "This Is How Much Anthropic and Cursor Spend on Amazon Web Services" — $2.66B AWS spend on $2.55B revenue through September 2025; costs scale linearly with revenue; estimates Google Cloud spend at minimum 50% of AWS

35. Anthropic — "Expanding Our Use of Google Cloud TPUs and Services" — Deal worth "tens of billions of dollars" for up to one million TPUs; over a gigawatt of capacity expected online in 2026

36. AI Planet X — "Cloud Spending Hits Anthropic" — Inference costs on Google and Amazon cloud ran 23% over plan; gross margin pushed toward ~40%; ~$5.2B burn in 2025

37. PM Insights — "Anthropic Approaches $7B Run Rate" — Over 1,200 employees by late 2025; plans to triple international workforce and quintuple applied AI team

38. Levels.fyi — Anthropic Salaries — Median yearly total compensation $471,808; highest reported $759,413 for Lead Software Engineer

39. Data Center Dynamics — "Anthropic Cloud Spend Expected to Reach $80B Through 2029" — $80B total cloud spend projected through 2029 across AWS, Google, and Microsoft; AWS takes up to 50% of gross profits; Google takes 20-30% of net revenue; $30B Microsoft Azure commitment

40. PYMNTS — "Anthropic Set to Pay Cloud Partners $80 Billion Through 2029" — Revenue sharing payouts: ~$1.3M in 2024, ~$360M in 2025, ~$1.9B in 2026, ~$6.4B in 2027 (The Information data)

41. Fierce Network — "What Does Anthropic's $50B Data Center Project Mean for Hyperscalers?" — $50B data center partnership with FluidStack; facilities in Texas and New York expected online throughout 2026

42. Sustainability Magazine — "Why Is Anthropic Pledging to Offset Its AI Energy Costs?" — Committed to covering 100% of grid upgrade costs and electricity price increases for communities hosting its data centers

43. Sacra — Anthropic Revenue, Valuation & Funding — Cash burn of $5.6B in 2024, ~$3B in 2025; projects cash flow positive in 2027, break-even in 2028

44. Axios — "Anthropic Raises $30B at $380B Valuation" — 300,000+ business customers; run-rate revenue of $14B; customers spending $100K+ annually grown 7x in past year

45. Built In — "2026 IPO Watchlist: OpenAI, SpaceX and Other Tech Giants" — Retained Wilson Sonsini (firm behind Google and LinkedIn IPOs) in December 2025 for IPO preparation

46. Newsweek — "AI Skeptic Ed Zitron Says Math on Data Centers Doesn't Add Up" — Claude Code subscribers consume $8-13.50 in compute per $1 paid; $200/month user burns $2,700+ in real costs; "Everyone using Claude Code right now is in an illusory environment"

47. Mercatus Center / George Mason University — "Private Sector Job Losses Dwarf Government Gains" — BLS Current Employment Statistics: government employment increased by 590,000 since January 2008 while private sector lost nearly 8 million jobs; government peaked at ~23 million in May 2010

48. Stanford HAI — The 2025 AI Index Report — Inference cost for GPT-3.5-level performance dropped over 280-fold between November 2022 and October 2024; driven by increasingly capable small models and open-weight alternatives

49. Epoch AI — "LLM Inference Prices Have Fallen Rapidly but Unequally Across Tasks" — Price declines ranging from 9x to 900x per year depending on benchmark; fastest trends (900x/year) start after January 2024; median rate increased from 50x to 200x per year when isolating post-2024 data

50. Bloomberg — "OpenAI Sees Better Margins on Business Sales, Report Says" — OpenAI compute margin reached 70% as of October 2025, up from 52% at end of 2024 and double the 35% rate in January 2024 (The Information data)

51. Les Barclays / Substack — "Who Captures the Value When AI Inference Becomes Cheap?" — Cost per query increasing for frontier models even as cost per token declines; models employing test-time scaling consume multiple tokens per problem; at current pace OpenAI's competitive advantage may dissipate

52. Foundation Capital — "NVIDIA's 'AI Factory' Bet" — Reasoning models require 20x more tokens and 150x more compute per query than traditional LLMs; enterprise applications projected to run hundreds of millions of agentic queries per month

53. iKangai — "The LLM Cost Paradox: How 'Cheaper' AI Models Are Breaking Budgets" — At maximum reasoning settings, models consume 10,000+ reasoning tokens internally for a 200-token answer; token consumption explosion leaving AI companies scrambling to stay profitable

54. AI Unfiltered — "The Inference Cost Paradox" — Enterprise AI spending surged 320% in 2025 despite per-token costs dropping 1,000x; organizations that budgeted for 20-30% increases found themselves scrambling by Q3; Jevons Paradox makes single-number AI budgets "fantasy"

55. LeadDigital — "A Token Economy" — Inference costs fell 1,000-fold but demand rose 10,000-fold; Jevons Paradox in action; Gartner predicts 40% of enterprise applications will feature AI agents by end of 2026, up from 5% in 2025

56. SaaStr — "Have AI Gross Margins Really Turned the Corner?" — Anthropic expects gross margin of 50% in 2025 and 77% by 2028; Cursor paying ~$650M/year to Anthropic on ~$500M revenue (negative 30% margin); Bessemer 2025 data shows AI startups averaging 25% gross margins; frontier models getting more expensive, not cheaper

57. California 2025-26 Budget Revenue Estimates — 67% of General Fund from personal income tax; capital gains = 16% of total tax liability (25% in boom years)

58. Wikipedia — "2008-2012 California Budget Crisis" — State revenue dropped from ~$100B to ~$85B during dot-com bust and again during Great Recession

59. Center on Budget and Policy Priorities — "Legacy of the Great Recession" — 8.7 million jobs lost in private sector during Great Recession

60. UC Berkeley Labor Center — "California Furloughs 2009" — 193,000 state workers furloughed; 2-3 unpaid days per month; 13.8% effective pay reduction

61. CalHR — Furlough Program Manual — Furlough hours banked as leave credits; paid out at separation at current base pay rate after COLA increases

62. Christian Science Monitor — "Government Jobs: A Recession Refuge?" (2009) — Public sector employment didn't drop during 1991 recession; didn't decline after dot-com bust until 2004; lagged private sector contractions by years

63. BLS Monthly Labor Review (2011) — "The Recession of 2007-2009" — State and local governments shed 68,000 jobs (0.3% decline) vs. millions in private sector

64. Governing — "State Budget Actions in Recessions" (2024) — FY2010: 26 states used layoffs, 22 states used furloughs as primary budget-balancing tool

65. California Legislative Analyst's Office — 2025-26 Budget Analysis — California has used furloughs in 9 fiscal years since 1992

66. PMC / Annals of the American Academy of Political and Social Science — "Safety Net Programs in the Great Recession" — Safety net caseloads rose from 276 million to 310 million recipients during the Great Recession; countercyclical demand expanded public sector workload

67. BLS — "Union Members — 2025" — Public sector union membership rate 32.9% vs. private sector 5.9%; local government workers at 37.8%

This analysis reflects the views of PenPublic and is informed by publicly available economic research and reporting. We encourage readers to review the cited sources and draw their own conclusions.

PenPublic — Your guide to public sector careers.

The career platform that converts America's workforce to public sector careers.

© 2025-2026 PenPublic