E115: The AI Search Wars: Google vs. Microsoft, Nordstream report, State of the Union

The condition of the prediction was met: the Russia–Ukraine war did in fact drag on for years after the podcast (Russia invaded in February 2022 and major combat and Western backing for Ukraine were still ongoing in late 2025).(en.wikipedia.org)

Olaf Scholz clearly ended up in serious domestic political trouble. His three‑party “traffic light” coalition became the most unpopular federal government in postwar Germany, with polls in 2023–2024 showing roughly 80–85% of Germans dissatisfied with its performance and his SPD sinking toward the mid‑teens in party support.(aa.com.tr) In November 2024, coalition infighting over economic policy led Scholz to sack his finance minister, triggering the collapse of the government and a confidence vote that he lost in December.(en.wikipedia.org) Early elections in February 2025 then produced a historic defeat for the SPD—its worst result since the 19th century at 16.4%—and Scholz lost the chancellorship to Friedrich Merz, who was elected by the Bundestag in May 2025.(en.wikipedia.org) Scholz’s fall from office after a snap election is strong evidence of “political trouble.”

A central driver of that trouble was economic pain linked to the loss of Russian gas and the war. Germany entered recession and became the only G7 economy to contract in 2023; economists and official analyses highlight the abrupt end of cheap Russian pipeline gas after the invasion and sanctions as a key cause of higher energy prices, inflation, and industrial weakness.(en.wikipedia.org) These energy‑price and growth problems remained a core domestic issue even into 2025.(reuters.com) Public-opinion work and political analysis of Scholz’s tenure repeatedly point to the energy crisis, inflation, and economic slump as major reasons for his collapsing approval ratings and the rise of protest parties like the AfD.(aa.com.tr)

Those economic problems are directly tied to Germany’s—and thus Scholz’s—decision to align with the U.S. and broader Western line on Russia after the invasion: backing sanctions that severed dependence on Russian gas, providing extensive financial and military support to Ukraine, and framing this as part of a joint Western strategy.(en.wikipedia.org) Domestic critics on both the left and right explicitly attacked this course, arguing that sanctions and arms deliveries to Ukraine—core elements of the U.S.-led policy—were damaging German economic interests and energy security; such themes featured prominently in AfD and other protest movements that gained as Scholz’s popularity collapsed.(en.wikipedia.org)

Other issues (immigration, internal coalition quarrels, climate and budget fights) also contributed, so the prediction slightly overstates how singularly U.S. alignment would be blamed. But in net: the war dragged on, the loss of Russian gas and war‑related economic pain became central domestic problems, Scholz’s pro‑Ukraine / pro‑sanctions stance as part of the U.S.-led camp was a visible part of that policy mix, and he ultimately suffered severe political consequences including the collapse of his government and loss of the chancellorship. That overall trajectory matches the prediction’s core claim, so it is best judged as right.

Evidence from Stanford’s 2025 AI Index shows that the cost of querying a GPT‑3.5‑level model fell from about $20 per million tokens in November 2022 to around $0.07 by October 2024—a more than 280× reduction in roughly 18 months, far exceeding the ~10× drop Chamath predicted would occur over “the next few years.” (therightstack.com) This comparison holds model quality constant (using MMLU to match GPT‑3.5‑level performance), so it is effectively measuring the cost of producing the “same output,” not of switching to weaker systems. (therightstack.com) The same report attributes these declines largely to hardware and efficiency gains: at the hardware level, compute costs declined about 30% annually while energy efficiency improved roughly 40% per year, matching Chamath’s emphasis on increasingly specialized silicon and improved energy efficiency as the drivers of cheaper inference. (shaswat.dev) Independent write‑ups of the AI Index and related analyses likewise highlight that inference prices for GPT‑3.5‑ or GPT‑4‑class capabilities have dropped by one to several orders of magnitude, with some models now matching GPT‑3.5/GPT‑4o performance at around $0.07 per million tokens, down from ~$20 in late 2022. (techopedia.com) While retail electricity prices themselves have not uniformly fallen, the effective energy and compute cost per inference has collapsed thanks to far more efficient accelerators (e.g., H100‑class GPUs delivering multi‑fold better inference performance and performance‑per‑watt than A100‑generation hardware) and much smaller, optimized models that achieve the same benchmark scores. (bestgpusforai.com) Taken together, by the mid‑2020s the per‑inference cost of producing ChatGPT‑style outputs is well below one‑tenth of the early‑2023 level, achieved via exactly the kind of hardware/efficiency curve he described, so his prediction is substantively correct (if anything, the cost decline has been far larger than he forecast).

Jason’s prediction has not come true as of November 30, 2025.

1. There is heavy IP/copyright litigation, but no shutdowns or “near‑death” events
Generative‑AI developers have indeed been hit with major lawsuits:

• The New York Times and other newspapers are suing OpenAI and Microsoft over training on news articles; a judge allowed core copyright claims to proceed, but did not order ChatGPT or Microsoft Copilot to be shut down or suspended.(apnews.com)
• The Authors Guild and groups of authors, as well as multiple newspapers, have parallel copyright suits against OpenAI and Microsoft; these are in discovery and briefing, not injunction/enforcement stages.(authorsguild.org)
• Visual artists are suing Stability AI, Midjourney, DeviantArt, and others; some claims were allowed to proceed, but again without any order halting their systems.(reuters.com)
• Record labels sued music‑generation startups Suno and Udio; this led to settlements and planned migration to licensed models and download limits, not service shutdown.(cnbc.com)

So Jason was right that “piracy/IP” litigation would arrive, but the effect has been lawsuits, settlements, and licensing deals—not stopping services “dead in their tracks.”

2. Courts have often trimmed or rejected the most extreme copyright theories
Key early cases have largely weakened the shutdown narrative:

• In the Getty Images v. Stability AI case in the UK, the High Court dismissed Getty’s core copyright claims, finding Stable Diffusion does not store or reproduce copyrighted works and rejecting secondary infringement; only narrow trademark issues (e.g., Getty watermarks in outputs) partially succeeded. No injunction barred Stable Diffusion’s operation.(washingtonpost.com)
• In the GitHub Copilot class action (Doe v. GitHub), the court has dismissed most of the 22 claims, including key DMCA copyright claims; only a couple of contract/license claims remain, and Copilot continues to operate.(theregister.com)
• Anthropic settled a major author lawsuit by agreeing to pay about $1.5 billion and destroy specific pirated book files, but the court found that the training itself was not illegal; Claude continues to run while Anthropic switches to lawful data sources.(apnews.com)
• A German court held that ChatGPT’s training on certain song lyrics violated German copyright and ordered damages, but did not order ChatGPT shut down.(theguardian.com)

These outcomes are economically painful and may change training practices, but they are a far cry from a Napster‑style injunction or a YouTube‑style existential “near‑shutdown.”

3. The flagship systems are expanding, not constrained like Napster/YouTube were

• OpenAI’s ChatGPT has exploded in usage: by 2025 it is handling on the order of 1–2.5 billion prompts per day, with hundreds of millions of weekly active users and projections of tens of billions in annual revenue, indicating ongoing expansion rather than a litigation‑induced freeze.(timesofindia.indiatimes.com)
• Microsoft has deeply integrated Copilot (formerly Bing Chat) into Windows, Microsoft 365, and the Windows 11 taskbar, treating it as a core platform feature rather than a legally precarious experiment.(en.wikipedia.org)
• OpenAI and others have signed broad licensing deals with major publishers (e.g., Future, Vox, News Corp, Financial Times, The Atlantic) specifically to reduce copyright risk while continuing to scale their models and products.(theverge.com)

In contrast, Napster was effectively forced offline by injunctions and could not continue its original service model; it shut down its network in 2001 and later went bankrupt.(en.wikipedia.org) YouTube, while not shut down, faced intense litigation (e.g., Viacom v. YouTube) that led to major compliance systems like Content ID and serious fears about its viability in the late 2000s.(en.wikipedia.org) Nothing comparable—no complete shutdown, no court‑ordered removal of the core model or service—has yet happened to ChatGPT, Microsoft Copilot, or similarly central generative‑AI products.

4. Some niche services have been reshaped, but not “stopped dead”
Music‑generation startups (Suno, Udio) and image models (e.g., Stability AI) have had to negotiate settlements, plan transitions to licensed datasets, or accept narrower operational constraints.(reuters.com) That is a meaningful level of legal friction, but still far from the prediction that generative AI as a class (particularly ChatGPT/Bing‑style systems) would experience a YouTube‑ or Napster‑level near‑death experience.

Conclusion
Jason correctly foresaw a wave of piracy/IP litigation against generative AI, but his stronger claim—that this litigation would halt or severely constrain major generative‑AI systems in a way comparable to Napster’s shutdown or YouTube’s early “near‑death” period—has not materialized. The flagship systems are larger and more deeply integrated into products than ever, with courts so far mostly trimming claims or steering outcomes toward damages, settlements, and licensing rather than existential injunctions.

Given the amount of time elapsed since the February 2023 prediction and the current legal and commercial landscape, the prediction is best classified as wrong rather than merely “too early to tell.”

As of November 30, 2025, major copyright battles around generative AI have produced money damages and some licensing frameworks, but not the kind of Content ID–style system Jason predicted.

• The largest generative‑AI copyright settlement to date is the $1.5B class‑action deal between authors/publishers and Anthropic. It creates a settlement fund that pays about $3,000 per copied book to rights holders who file claims, compensating past use of pirated copies, but it does not create an ongoing system to detect future use of individual works in training or outputs, nor a standing dashboard where rightsholders can register content and choose monetization options as on YouTube.(reuters.com)
• Other headline lawsuits against generative‑AI companies – e.g., the New York Times and other newspapers vs. OpenAI and Microsoft – remain active with core copyright claims still moving toward trial, so no settlements from them have created such an infrastructure either.(apnews.com)
• Visual‑art cases (artists vs. Stability AI/Midjourney/DeviantArt; Getty Images vs. Stability AI) have led to partial dismissals and, in the UK Getty case, a ruling largely favoring Stability AI on core copyright issues, not to a settlement that mandates a universal identification/claims system.(insideglobaltech.com)
• In music, record‑label lawsuits against AI music startups Suno and Udio have recently been settled through licensing deals that allow new licensed models and impose download restrictions. These are standard catalog licenses with opt‑in by label artists; they do not create a generalized mechanism for all rightsholders to see when their works were used in training and to claim ongoing per‑use compensation across AI platforms.(reuters.com)
• Separately from litigation, some companies have built limited compensation schemes—e.g., Shutterstock’s Contributor Fund shares revenue with contributors whose content is in licensed training datasets, and Adobe trains Firefly on licensed/stock content while compensating contributors—but these are voluntary business arrangements, not outcomes of major copyright settlements, and they still don’t let any rightsholder upload a work, have the system detect training or output uses across the industry, and choose to monetize or block them in a Content ID–like way.(submit.shutterstock.com)
• Tools like Have I Been Trained? do let creators search certain public datasets such as LAION‑5B to see whether their images appeared in training data, but this is an independent activist project, not a settlement‑mandated system, and it offers opt‑out/advocacy rather than integrated, ongoing monetary compensation.(makeuseof.com)

Because no major generative‑AI copyright settlement has yet produced a YouTube‑style, industry‑level system that (a) identifies when specific copyrighted works were used in training or outputs and (b) lets rightsholders claim ongoing compensation based on that usage, Jason’s specific prediction has not come true to date, even though some partial, narrower analogues (class‑action funds, catalog licenses, contributor funds) have emerged.

Evidence so far is mixed and timing-dependent.

What the prediction claimed
Chamath argued that as LLM-based AI is widely deployed in search and adjacent products, it will:

1. Be strongly deflationary.
2. Drive aggregate industry revenues and profit pools down, including Google’s, unless incumbents cannibalize their existing models.

What has actually happened (through late 2025)

1. Google’s search and ad revenues are still rising in absolute terms.

   • Alphabet’s Google Search & Other revenue grew from about $162.5B in 2022 to $175.0B in 2023 and $198.1B in 2024. Total Google advertising revenue rose from $237.9B in 2023 to $264.6B in 2024, and total company revenue from $307.4B to $350.0B. (reportify.ai)
   • In 2025, Alphabet has reported multiple record quarters, including over $100B in quarterly revenue and roughly 16% year‑over‑year growth, with nearly $35B in quarterly profit. (apnews.com)
   • Google’s search revenue specifically was reported up ~10% year‑over‑year in 2025, not down. (mediaweek.com.au)
     → This contradicts a near‑term drop in Google’s revenue/profit pool.

2. The broader search and digital ad markets are also still growing.

   • Global search advertising revenue reached about $168.9B in 2024 and is forecast to keep growing strongly. (grandviewresearch.com)
   • Overall digital ad spending was about $614B in 2022 and ~$600–740B range in 2024, with projections for continued growth through the late 2020s. (theseoproject.org)
     → At the aggregate market level, revenues are not yet falling; they’re growing, albeit with some deceleration.

3. There is clear deflationary pressure and profit compression for parts of the ecosystem.

   • Google’s AI Overviews/AI search features are measurably reducing click‑through rates and traffic to publishers, with reports of 18–64% drops in organic visits and large CTR declines where AI answers appear. (searchinfluence.com)
   • Multiple reports document steep traffic and revenue declines at news and content sites (e.g., 30–40%+ traffic drops, rising zero‑click searches, and sharp falls in affiliate revenue), attributed substantially to AI summaries and Overviews. (ainvest.com)
   • Analysts also note Google’s ad revenue growth is slowing and that AI features are harder to monetize, while infrastructure capex for AI is surging, pressuring margins. (searchengineland.com)
     → This supports the directional claim that AI in search is deflationary for many incumbents and compresses parts of the profit pool.

4. Incumbent self‑cannibalization is happening, which muddies the counterfactual.

   • Google has, in effect, done what Chamath advised: aggressively integrating Gemini, AI Overviews, and an AI “mode” directly into core search, then adding ads into those AI experiences. (techcrunch.com)
   • Because Google did choose to partially cannibalize its old SERP model, we can’t directly observe the alternative world where it refused to do so and then lost revenue to competitors, as his conditional prediction envisioned.

5. Timing: “as AI becomes widely deployed” is not clearly satisfied yet.

   • LLM‑driven search is still in early deployment: AI Overviews appear on roughly 30–40% of searches, AI “modes” have ~100M MAUs in a few markets, and AI‑driven search ads are only about $1B in U.S. spend in 2025, projected to grow sharply by 2029. (websiteplanet.com)
   • We’re mid‑transition rather than in a mature, fully AI‑dominated search landscape, so his long‑run claim about aggregate revenue and profit shrinkage may simply not have had time to play out.

Why the verdict is “inconclusive” rather than clearly right or wrong

• Against “right”: So far, the aggregate revenue and profit pools for search/digital ads and for Google itself have grown, not shrunk, which contradicts the narrow reading of his claim if judged strictly on 2023–2025 data. (alphaminr.com)
• Against “wrong”: His structural thesis—AI search as a deflationary force that erodes parts of the existing web/media profit pool and pressures margins—is strongly supported by the observed collapse in traffic and revenue for many publishers and the mounting difficulty of monetizing AI search. (ainvest.com)
• Too early: Because LLM‑based search is only partially deployed and AI ad formats are nascent, it’s plausible that aggregate revenues could flatten or fall later this decade, but current data can’t confirm or rule that out.

Given these factors, the fairest assessment as of November 30, 2025 is that Chamath’s prediction is directionally supported in terms of deflationary pressure and ecosystem damage, but its strong quantitative claim about aggregate revenues and profit pools going down has not yet materialized. Therefore the outcome is best labeled “inconclusive (too early)” rather than definitively right or wrong.

As of late 2025, the specific outcome Friedberg described has not occurred, but the prediction doesn’t have a clear time horizon, so it can’t be definitively scored as right or wrong yet.

What the prediction claimed
He argued that as AI answer engines mature, around 95% of low–value, marginal content sites in areas like self‑help and basic how‑to advice would “go away,” because users would instead get aggregated, synthesized answers directly from AI interfaces.

What we see today (Nov 2025)

1. AI answer engines have clearly “matured” and become mainstream.

   • Google’s AI Overviews (successor to SGE) now appear in over 50% of Google search results, up from ~25% just ten months earlier, and are heavily skewed toward informational queries. (en.wikipedia.org)
   • Consumer use of AI tools for search‑like tasks (ChatGPT, Gemini, Perplexity, Copilot, etc.) has risen to ~38% of users in some markets, with ChatGPT alone counted among the most visited sites globally. (en.wikipedia.org)
     This broadly matches the “AI answer engines” environment he envisioned.

2. AI and Google updates have hurt many low‑value content sites, but not wiped out ~95% of them.

   • Google’s March 2024 “helpful content” and spam updates targeted low‑quality, unoriginal, and mass‑produced AI content, aiming to remove around 40% of such material from search results; later analysis suggests roughly a 45% reduction of low‑quality/unoriginal content in SERPs. (samblogs.com)
   • A study of Google’s March 2024 update found that 837 of 49,345 monitored sites were completely deindexed in early stages—hundreds of sites, not anything close to 95% of a whole niche. (searchenginejournal.com)
   • Analyses of deindexed sites show many had very high proportions of AI‑generated content and aggressive ads—again indicating a culling of some spammy/low‑value sites rather than near‑total extinction. (originality.ai)

3. AI summaries are clearly siphoning traffic from publishers, but sites largely still exist.

   • SimilarWeb data show that after AI Overviews rolled out, news‑related zero‑click searches on Google rose from 56% to 69%, and organic traffic to news sites fell from over 2.3B monthly visits to under 1.7B between mid‑2024 and May 2025. (nypost.com)
   • Reports and complaints by publisher groups in the US and EU describe substantial traffic and revenue declines, and some individual businesses (like Chegg) explicitly blame AI search features for subscriber and traffic loss, leading to layoffs and lawsuits. (reuters.com)
     These are serious impacts, but they show declining traffic and monetization, not that most how‑to / self‑help sites have vanished.

4. The overall number of websites is roughly stable or growing, not collapsing by 95%.

   • Multiple industry summaries estimate roughly 1.1–1.2 billion websites worldwide in 2025, with around 190–260 million active sites—similar active‑site levels to 2020–2023. (grabon.com)
   • The web continues to add ~250,000 new websites per day, and tens of millions of new domains yearly. (grabon.com)
     There is no sign that 95% of self‑help/how‑to publishers have disappeared; if anything, generative AI has increased the supply of low‑effort content in many niches. Economic pressure and deindexing are real, but they fall well short of the near‑total wipe‑out he envisioned.

Assessment

• On the mechanism, he was directionally right: AI answer engines and Google’s AI Overviews are now major interfaces for informational queries and are measurably reducing click‑through to many content sites, especially low‑value or duplicative ones. (en.wikipedia.org)
• On the magnitude and endpoint—“95% of these content publishers…go away” as a result—available data don’t support this. The total number of websites has not collapsed, and while many low‑value sites have lost visibility or revenue and some have been deindexed, that is nowhere near a 95% disappearance of self‑help/how‑to style sites.

Because Friedberg didn’t specify a time frame (e.g., “by 2025” or “within five years”), we can’t yet say the prediction is definitively false; the structural transition he described could continue for the rest of the decade. As of late 2025, though, the core quantitative claim (≈95% of such sites disappearing) has not come true, and the evidence so far points to substantial attrition and consolidation rather than near‑total extinction—hence the status **“inconclusive (too early)” rather than clearly right or clearly wrong.

The prediction is framed as a conditional: if Microsoft were to take roughly 5–6 percentage points of search share from Google via AI, then major U.S. antitrust efforts against Google and other tech giants would effectively lose credibility and no major adverse antitrust remedies would be imposed.

On the antecedent (market-share shift):

• In the U.S., Bing’s total search share was about 6.4% in Q2 2023 with Google around 89%.
• By early 2025, estimates put Google at roughly 88% and Bing at 7–8.5% of U.S. search, i.e., a gain of only about 1–2 percentage points, not the 5–6 points Chamath specified. (wpshout.com)
• Globally, Bing’s share rose from about 3.0% in 2023 to ~3.9–4.1% in 2025, again roughly a 1 point gain, not 5–6. (backlinko.com)
  So the key condition — Microsoft taking 500–600 basis points of search share from Google — did not occur.

On the consequent (antitrust becoming DOA):

• The U.S. Department of Justice (DOJ), not the FTC, won landmark cases finding Google to have illegal monopolies in both general search and in key digital ad-tech markets, with courts ordering behavioral remedies (data-sharing with rivals, limits on exclusivity) and opening the door to potential structural remedies in ad tech. (theguardian.com)
• The FTC has remained active against Big Tech: it brought a major antitrust suit alleging Amazon illegally maintains an online retail monopoly (trial set for 2027) and secured a record $2.5 billion settlement over deceptive Prime sign‑up and cancellation practices. (en.wikipedia.org) It also pursued a breakup case against Meta over Instagram/WhatsApp, which it ultimately lost in November 2025, but the case went all the way through trial and a merits ruling. (businessinsider.com)

These outcomes show that antitrust enforcement against Big Tech has by no means disappeared, though the FTC’s litigation record is mixed. However, because the specific trigger condition (a 5–6 point Bing share gain) never materialized, we cannot directly test Chamath’s asserted causal relationship between such a shift and the demise of credible antitrust action.

Therefore, strictly on the normalized conditional prediction — if Microsoft gains ~5–6 points of search share, then major antitrust remedies become DOA — the real world has not satisfied the "if" clause, so whether the "then" clause would have followed is inherently unknowable from observed data. That makes the prediction ambiguous, not cleanly right or wrong.

As of 30 November 2025, the United States federal government has not defaulted on its debt obligations in the usual sense of failing to pay interest or principal when due. There have been repeated debt-ceiling standoffs (e.g., in 2023), but these were resolved via legislation before any formal default occurred; ratings agencies have sometimes responded with downgrades or warnings, but not in response to an actual missed payment on Treasuries.

Friedberg’s prediction is explicitly about an eventual default under certain long‑run policy conditions (no entitlement cuts, continued servicing of existing debt plus interest, and roughly unchanged tax rates). Because he gave no time horizon and framed this as something that must happen "at some point" in the future if trends continue, the claim cannot yet be evaluated: the relevant future period is still unfolding, and no disconfirming default event (or decisive policy reversal that clearly removes the risk) has occurred.

Given that (1) the U.S. has not defaulted, and (2) the prediction concerns an unspecified future date, it is too early to judge whether the prediction will ultimately be right or wrong. Therefore the appropriate status is inconclusive (too early).

The prediction is explicitly about what could happen "if" the U.S. in the coming decades (a multi‑decade timeframe) (a) reduces energy costs by 50–75% and (b) increases energy production capacity by 10–20x, and then potentially grows out of its long‑term debt/entitlement problem.

As of November 30, 2025:

1. Timeframe: Only about 2 years and 9 months have passed since the prediction (February 11, 2023). The relevant horizon is “decades,” so the specified period has not elapsed.
2. Condition not met: The U.S. has not experienced a 50–75% nationwide reduction in energy costs nor a 10–20x increase in total energy production capacity; instead, energy capacity has grown incrementally, and prices have fluctuated but not collapsed by that magnitude on a sustained, economy‑wide basis. This means we have not reached the scenario under which Friedberg claims growth could solve the debt/entitlement problem.
3. Outcome untestable: Because (a) the condition has not occurred, and (b) the multi‑decade evaluation window has not passed, we cannot judge whether his conditional claim about long‑run debt dynamics would have been correct.

Therefore, the prediction’s correctness cannot yet be evaluated; it remains a conditional, long‑term scenario, and it is too early (and the condition has not even been triggered) to classify it as right or wrong.

The prediction explicitly concerns outcomes over a 20–30 year horizon starting from 2023:

• Time window for 10x U.S. energy production capacity: roughly 2043–2053.
• Time window for a potential disruptive breakdown in at least one of entitlements, taxes, or debt if that 10x increase does not occur: also centered around that same 2043–2053 period.

As of today (November 30, 2025), we are only about 2–3 years into that 20–30 year window. It is far too early to know:

1. Whether U.S. energy production capacity will in fact increase by ~10x over the full 20–30 year period.
2. Whether, if it does not increase by that amount, a clearly “ugly” breakdown in entitlements, tax regime, or debt sustainability will occur by around 2043–2053.

While we can observe partial trends in U.S. energy production and fiscal policy today, those cannot yet confirm or falsify a contingent, long-horizon scenario framed explicitly around 2043–2053. Therefore the accuracy of the prediction cannot presently be evaluated.

As of November 30, 2025, there is still no commercially viable fusion power plant (magnetic confinement, inertial confinement, or other non-solar fusion) operating at scale on the grid. All fusion projects remain in the R&D / demonstration / pilot phase, with no utility-scale commercial plants selling power into markets.

At the same time, grid‑scale energy storage has grown substantially (especially lithium‑ion battery projects, plus pumped hydro and some emerging long‑duration storage), and is already being deployed alongside large amounts of wind and solar. However, this deployment has not yet clearly solved the problem in Chamath’s framing: enabling “abundant zero(-marginal‑)cost energy” in a way that removes storage as the “real technical bottleneck.” Storage is still widely seen as a key constraint to very high (>80–90%) renewable penetration, and no consensus exists that we have fully “figured out” grid‑scale, long‑duration, highly scalable storage to that degree.

Because Chamath’s prediction is explicitly about which comes first in the future—

"we'll have [the storage solution] well before different forms of fusion are commercializable"

—the relevant question is: by the time fusion does become commercially viable at scale, will grid‑scale storage sufficient for abundant near‑zero‑marginal‑cost renewables have already been developed and deployed?

As of late 2025:

• Fusion is still not commercially viable.
• Storage is improving and scaling, but it is not yet clear that we’ve “figured out” the bottleneck to the level implied.
• The sequence Chamath predicts (storage solution clearly arriving well before commercial fusion) has not yet been definitively resolved, because neither side of the comparison (commercial fusion vs. clearly sufficient long‑duration storage) has hit its decisive milestone.

Given that the core claim is about future ordering rather than a fixed date, and we have not yet reached the point where either condition is unambiguously satisfied, the prediction cannot yet be judged as right or wrong.

Therefore, the status is inconclusive (too early).