E106: SBF's media strategy, FTX culpability, ChatGPT, SaaS slowdown & more

There is evidence consistent with parts of Chamath’s claim, but it is not strong or clean enough to say the prediction is clearly right or clearly wrong.

Evidence that large grifts have continued and some elite coverage remained sympathetic

• In 2023 the Associated Press and others documented what they called “the greatest grift in U.S. history”: an estimated ~$280B in stolen and another ~$120B in wasted or misspent U.S. COVID‑19 relief funds, roughly 10% of all aid disbursed. That’s a huge, system‑scale grift persisting into and being fully recognized after 2022, matching the prediction’s claim that large frauds would continue in the U.S. landscape. (wusf.org)
• In 2023 Michael Lewis published Going Infinite about Sam Bankman‑Fried. Major reviews in outlets like the Washington Post and New York Times described the book as “stubbornly credulous” and overly sympathetic to SBF, saying Lewis seemed “snowed” and spent more time second‑guessing FTX’s bankruptcy leadership than drilling into the fraud. Critics saw this as a prominent media/elite figure effectively giving SBF a partial “hall pass” even after the collapse. (en.wikipedia.org) This lines up, at least rhetorically, with Chamath’s complaint that some in‑group figures get unusually forgiving narratives.

Evidence cutting against the strong form of the media‑protection claim

• Mainstream U.S. outlets also did extensive critical coverage of SBF and FTX once cracks appeared, and prominently reported his 2023 federal trial, conviction on seven fraud and conspiracy counts, and 25‑year sentence. (en.wikipedia.org) That undercuts a simple story where “mainstream media” reliably shield in‑group fraudsters from accountability.
• For large post‑2022 fraud stories like COVID‑relief fraud, mainstream coverage has mostly focused on exposing and quantifying the abuse (e.g., AP’s “Great Grift” investigation and follow‑on coverage in outlets such as the Philadelphia Inquirer and Fortune), not on protecting the perpetrators. (inquirer.com) The drivers were weak government controls and speed of disbursement, not clearly the kind of media in‑group deference Chamath describes.

Why the prediction can’t be cleanly scored

• The prediction bundles two elements: (1) a mechanism (“mainstream media protect ideological/class in‑group figures and don’t admit errors”) and (2) an outcome (“similar grifts will continue”).
• The outcome is so broad and open‑ended that it is almost guaranteed to be true in some sense; large frauds and grifts do keep happening. But connecting specific post‑2022 frauds causally to media protection of in‑group elites is highly subjective and not something the available evidence can decisively prove or falsify.
• For every example someone could cite as confirmation (e.g., credulous SBF narratives in parts of elite media), another person can point to strong investigative and prosecutorial work as counter‑evidence that the system, including much of the press, ultimately punished rather than protected the grifters.

Because of this mix of partially supportive anecdotes, substantial counter‑evidence, and the broad, value‑laden nature of the claim, the prediction’s truth value is fundamentally interpretive rather than empirically crisp, so it is best classified as ambiguous.

Evidence strongly supports the direction of Friedberg’s prediction (creators disrupting traditional media and journalism), but not clearly the specific claim that independent/creator media already constitute the majority of media volume consumed.

Supporting evidence for the disruption / rise of creators

• The global creator economy is now a very large, fast‑growing market: estimates put it around $200–250B in 2024, projected to exceed $1T by the early 2030s, with individual content creators responsible for roughly 60% of that revenue. (grandviewresearch.com)
• Ad revenue on creator-driven platforms (YouTube, TikTok, etc.) is forecast in 2025 to overtake that of traditional media (TV, print, radio, cinema) for the first time, a major shift in where money and attention flow. (businessinsider.com)
• Streaming has surpassed broadcast + cable in U.S. TV viewing share, with YouTube alone capturing around 12.5% of all TV viewing time—driven largely by user‑generated and creator content—and social video platforms now accounting for about 20% of all TV viewing. (technology.org)
• For news specifically, a 2025 Reuters Institute report finds that, at least around the U.S. presidential inauguration, more Americans said they got news from social and video networks than from TV, news sites or apps, and it highlights online personalities (Rogan, Carlson, etc.) as increasingly central news sources. (reuters.com)
• Traditional journalism shows clear signs of disruption: Vice Media’s bankruptcy and layoffs, the shutdown of Vice.com news publishing and BuzzFeed News, repeated large layoffs at outlets like CNET and major newspapers such as The Washington Post, and ongoing declines in legacy news economics. (cnbc.com)

Where the prediction is not clearly fulfilled

• Pew finds about 1 in 5 U.S. adults regularly get news from “news influencers” (independent social‑media personalities, mostly unaffiliated with news orgs), rising to roughly 37–38% among adults under 30. While a majority of these influencers are indeed independent (77% have no news‑org affiliation), this is still far from a majority of the overall population’s news consumption. (pewresearch.org)
• On TikTok, among people who use it for news, influencers/celebrities and other non‑journalist accounts are as or more important than news outlets—but that’s one platform and subset of users, not total media consumption. (pewresearch.org)
• In some countries (e.g., Australia), social media has just overtaken legacy online sources as a main news source, but TV remains the single biggest source, and traditional outlets still command large audiences. (theguardian.com)
• Even on TV, where creator platforms are surging, social/creator video accounts for roughly a fifth of viewing time, not a clear majority of all media usage across devices. (tvtechnology.com)

Why the result is “ambiguous”

• There is strong, measurable evidence that traditional journalism/press is being heavily disrupted by creator-style media and that creators are capturing a rapidly growing share of attention and revenue.
• However, whether independent journalists and creator-type media now constitute the majority of all media volume consumed is not cleanly measurable and is not clearly supported by available data. In most broad, population‑level metrics we can see (news sources, TV viewing, etc.), creator content is large and growing but not unambiguously a majority.
• The timeframe (“in the coming years”) is also vague; by late 2025, many trend elements look directionally correct, but the strong form of the prediction (clear majority dominance) can’t be definitively confirmed or falsified.

Because the disruption part is clearly happening but the “majority of media volume” claim is not clearly met or refuted with current data, the fairest evaluation is ambiguous, rather than fully right or fully wrong.

Key elements of Chamath’s prediction have largely played out:

1. Consolidation as de‑facto “ruler for life” in 2022
   In October 2022, at the Chinese Communist Party’s 20th Party Congress, Xi Jinping secured a precedent‑breaking third term as party leader, stacked the Politburo Standing Committee with loyalists, and left no clear successor, leading major outlets and analysts to describe him as China’s most powerful leader since Mao and a potential ruler for life.(washingtonpost.com) Subsequent scholarship and events have routinely referred to Xi as China’s “ruler for life,” including works by Willy Lam and academic events explicitly using that phrase.(catalog.k8s.libraries.psu.edu) This aligns closely with Chamath’s claim that 2022 would be the first year he is “essentially really ruler for life.”

2. 2022 as a year of major political “winning”
   The same 20th Party Congress and ensuing leadership reshuffle were widely viewed as a huge personal victory for Xi, cementing his dominance over the party and state.(washingtonpost.com) While the label “biggest political winner globally” is subjective—others might argue for figures like Volodymyr Zelensky in 2022—it is not contradicted by events; if anything, much commentary supports the view that Xi was one of the year’s principal political winners due to his unprecedented consolidation of power.

3. Expanded power playing out domestically and internationally after 2022
   After the podcast:

   • In March 2023, Xi was confirmed for a third term as China’s president, reinforcing his consolidated authority.(outlookindia.com)
   • Domestically, analyses note continued centralization around Xi, extensive anti‑corruption and faction‑purging campaigns, and tighter ideological and surveillance control, all framed around his personal rule.(abc.net.au)
   • Internationally, Xi has advanced initiatives like the Global Development Initiative and Global Security Initiative, deepened ties with Russia and the Global South, and pursued a more assertive foreign policy, using China’s influence in events such as the Iran–Saudi rapprochement and continued military and diplomatic pressure around Taiwan and the South China Sea.(outlookindia.com)

These developments match Chamath’s core forecast: Xi used 2022 to lock in an effectively open‑ended tenure and that enhanced power has since been expressed in both domestic governance and international strategy. The only clearly subjective piece—“biggest political winner globally”—cannot be decisively falsified and is broadly consistent with how many observers characterized his 2022 power consolidation. Therefore, on balance, the prediction is best judged as right.

Chamath’s core claim was that in 2023 Xi Jinping would (a) tighten control at home and (b) act aggressively abroad by leveraging China’s dominance over critical supply chains. Both elements are borne out by events.

1. Leveraging control of critical supply chains (the key mechanism)

• China has overwhelming shares of production or processing in several critical minerals. For example, a CSIS analysis notes that by 2024 China produced about 98% of the world’s low‑purity gallium and dominated global germanium and antimony output, giving Beijing unusual leverage over these supply chains.(csis.org)
• On July 3, 2023, China’s Ministry of Commerce and customs authorities announced that exports of gallium and germanium and related compounds would require government licenses starting August 1, 2023, under the Export Control Law and other security legislation.(mayerbrown.com) These metals are critical for semiconductors, EVs, telecoms, and defense systems, so such controls directly weaponized supply-chain dominance.
• After the curbs took effect, customs data showed China exported zero gallium and germanium products in August 2023, a sharp drop from July, demonstrating their potential to choke off supply.(cnbc.com)
• In October 2023, China imposed formal export controls on several categories of high‑purity graphite used in EV batteries and other advanced technologies; exporters now need licenses, and foreign EV makers were warned of likely supply disruption.(chinastrategy.org) Given China’s dominant share of natural and processed graphite for battery anodes, this again translated supply-chain control into geopolitical leverage.
• Contemporary analyses explicitly describe these moves as part of Beijing’s emerging toolkit of economic coercion / export weaponization in response to Western semiconductor controls, matching Chamath’s thesis that Xi would exploit control over “critical supply chain[s] that the Western world needs.”(csis.org)

2. Running “roughshod” domestically: expanded security state and party control

• On April 26, 2023, the NPC Standing Committee passed a sweeping revision of China’s Counter‑Espionage Law, effective July 1, 2023. The law massively broadens “espionage” to cover “all documents, data, materials, and items related to national security and interests” and grants security agencies wide powers to access data, search property, and restrict travel—changes that legal and business analysts say significantly increase risks for foreign firms, journalists, and NGOs and enhance the security apparatus’ discretion.(loc.gov)
• Also on June 28, 2023, China passed a new Foreign Relations Law, in force from July 1, which codifies CCP leadership over all foreign policy and explicitly provides a legal basis for “countermeasures” and “restrictive measures” against foreign states, embedding a confrontational, sovereignty‑maximalist posture in statute.(en.wikipedia.org)
  These moves fit the idea of Xi consolidating and using domestic power in a heavy‑handed, security‑driven way.

3. International behavior: mix of aggression and pragmatism, but with notable hard‑edged moves

• The 2023 Chinese balloon incident, in which a large Chinese balloon traversed U.S. and Canadian airspace before being shot down off South Carolina on February 4, 2023, became a major diplomatic confrontation, worsening already poor U.S.–China relations and reinforcing Western perceptions of Chinese assertiveness and disregard for others’ airspace.(en.wikipedia.org)
• The 2023 export controls on gallium, germanium, and graphite were widely read as retaliatory measures against U.S. and allied chip export curbs, underlining a willingness to weaponize economic interdependence rather than play a purely cooperative role in global trade.(globalpolicywatch.com)
• At the same time, 2023 also saw more pragmatic or stabilizing diplomacy, such as China brokering the Saudi–Iran normalization agreement announced on March 10, 2023, which was welcomed internationally,(en.wikipedia.org) and some signs of toning down the most flamboyant “wolf‑warrior” rhetoric (for example, sidelining high‑profile hardliners like Zhao Lijian early in the year).(theguardian.com) So the style of diplomacy was not uniformly more abrasive than in 2020–2022.

4. Why this still counts as the prediction being basically right

• Chamath’s specific mechanism—that Xi would leverage China’s control over critical supply chains to project dominant, often coercive influence abroad in 2023—is clearly borne out by the new export controls on gallium, germanium, and graphite and the way they were received by affected countries and industry.(mayerbrown.com)
• Domestically, Xi continued to centralize authority and expand the security state via the revised Counter‑Espionage Law and related national‑security legislation, aligning with the idea of “running roughshod” at home.(loc.gov)
• While Beijing also engaged in selective stabilization efforts and high‑profile mediation, these do not negate the fact that 2023 is the year China visibly tested and deployed supply‑chain leverage as a strategic tool, which is the heart of the forecast.

Given that the central, falsifiable parts of the prediction—use of supply‑chain dominance for aggressive leverage and continued hardline domestic control in 2023—did occur, the prediction is best judged as right, albeit with the nuance that China’s diplomacy also included some moderating and bridge‑building elements alongside the coercive ones.

Key elements of Sacks’s prediction broadly matched what happened in 2023:

1. Xi Jinping keeps or strengthens his grip on power
   Xi was unanimously reelected state president for an unprecedented third term on March 10, 2023, after the 2018 constitutional change removing presidential term limits, and analyses describe his administration as “rock solid” with a top leadership dominated by loyalists and no visible internal challenge to his authority. (en.wikipedia.org) This aligns with the claim that Xi would not “lose his grip in any way.”

2. Worsening conditions for Chinese citizens and private sectors (billionaires, tech, real estate)
   Billionaires/tech: China lost 229 billionaires from the Hurun Global Rich List 2023, more than half of all those who fell off the list worldwide, with the decline explicitly linked to Beijing’s crackdown on major tech companies and other headwinds. (dawn.com)
   Real estate/households: China’s property downturn deepened: property investment fell sharply in 2023 and new construction starts plunged; developers such as Country Garden missed bond payments, and the housing slump eroded household wealth and confidence. (euromonitor.com) Household sentiment, youth employment and equity markets were weak, with research noting high youth unemployment (above 20%), falling property prices, and a CSI 300 stock index down about 35% from mid‑2021 to end‑2023, all weighing on consumer confidence. (tspr.org) This is consistent with things getting “worse for everybody else in China,” especially the private sector and middle class.

3. 2023 not being a “good year” for China’s economy
   While official GDP growth around 5.2–5.4% in 2023 met or slightly exceeded Beijing’s headline target, the year was widely characterized by economists as a disappointing, weaker‑than‑expected post‑COVID recovery: soft consumption, weak exports, a deepening property crisis, and rising deflation pressures. (thenews.com.pk) Analysis from banks and research houses emphasized that deflation, property-sector contraction, and fragile confidence made 2023 a structurally poor year for China’s economy despite the headline growth figure, which matches Sacks’s directional call that it would “not be a good year.”

4. Potential contagion from China
   Sacks framed contagion as a risk (“I think there could be contagion”), not as a certainty. China’s real estate turmoil did in fact generate financial‑stability worries: liquidity strains at major developers, missed bond payments, and growing non‑performing loan risks for banks, with analysts explicitly warning that property stress could spill over into broader financial markets and the global economy. (euromonitor.com) While this did not culminate in a 2008‑style global crisis, the emergence of notable spillover risk is directionally consistent with his warning.

Because Xi’s political power remained firmly intact while 2023 turned out to be a weak, confidence‑sapping year for China’s economy—especially for real estate, private business, and the wealthy—and because the kind of financial‑sector and contagion risks he highlighted did materialize (even if not catastrophically), the overall prediction is best judged as right on direction and substance.

Evidence after 2023 does not support the prediction that a "restarted" Chinese growth engine would meaningfully drive U.S. growth or push U.S. inflation higher.

1. China’s post‑reopening growth was middling and structurally weak, not a restored “engine.”
After dropping zero‑Covid, China’s GDP did rebound to about 5.2% in 2023 and around 5% in the first half of 2024, helped by reopening and public investment. But the IMF and others emphasize that this growth came with significant slack, very low inflation, and continued property‑sector stress, and they project a steady slowdown to near 3% later in the decade.(imf.org) This is far from a return to the pre‑2015 “high‑octane” growth engine that reliably pulled the global economy.

2. U.S. inflation fell sharply despite China’s reopening, and China increasingly acted as a disinflationary force.
U.S. CPI inflation dropped from about 8.0% in 2022 to roughly 4.1% in 2023 and 2.9% in 2024, according to BLS-based series.(bls.gov) Over this same period, China slipped into producer‑price and intermittent consumer‑price deflation; economists at PIMCO and others noted that weaker Chinese demand, falling factory‑gate prices, and a softer yuan were lowering the prices of Chinese goods abroad and dampening global commodity demand, explicitly describing this as helpful to Western central banks battling high inflation.(cnbc.com) That is the opposite of a renewed Chinese boom significantly re‑inflating U.S. prices.

3. U.S. growth was driven mainly by domestic demand, not a China‑led external boost.
U.S. real GDP grew about 2.5% in 2023 and 2.8% in 2024, with BEA attributing the gains primarily to domestic consumer spending, business investment, and government outlays; exports were a secondary contributor.(bea.gov) IMF global outlooks through 2024–25 describe world growth as “modest” by historical standards, with the U.S. surprisingly strong while China’s property problems and weak confidence are a drag on prospects rather than a major positive spillover.(apnews.com) That pattern doesn’t match a story in which revived Chinese growth significantly boosts U.S. GDP.

4. Net assessment versus the prediction.

• “Chinese growth engine will resume” – only partly: China did rebound numerically, but the recovery was fragile, quickly confronted deflation and property‑sector crisis, and is widely characterized as a structurally slowing economy rather than a restored global engine.(imf.org)
• “…and this renewed Chinese growth will have significant effects on U.S. economic growth and U.S. inflation” – not borne out: U.S. inflation and growth paths are better explained by domestic policy and demand; China’s contribution appears modest and in inflation’s case largely disinflationary, not a sizable new inflation or growth impulse.

Given these outcomes, the specific forecast that China’s reopening would restart a powerful growth engine that then has significant effects on U.S. growth and inflation is best judged wrong.

The prediction has effectively come true by late 2025.

1. “Tens of thousands of startups” around generative AI
Multiple market analyses now estimate tens of thousands of generative‑AI firms worldwide. Several 2024–2025 industry reports put the number of generative‑AI startups/companies at roughly 67,000+ globally, up from about 50,000 at the end of 2023, with projections that it could approach 100,000 if current trends continue. (seo.goover.ai) Broader AI tracking from Tracxn similarly finds over 60,000 AI companies as of 2024, consistent with the “tens of thousands” order of magnitude. (artsmart.ai) This matches Friedberg’s qualitative claim that roughly 100,000 startups would emerge around this wave of technology.

2. Generative AI as the focal point of the next hype/bubble cycle
Venture data show AI—driven heavily by generative AI—became the clear center of the funding cycle after 2022:

• AI’s share of global startup funding jumped from ~13% in 2022 to about 33% by late 2024, even as overall VC volumes fell, with global generative‑AI funding alone surging from under $1B years earlier to around $40–45B in 2024. (link.springer.com)
• By the first half of 2025, AI startups accounted for about 53% of global VC funding and 64% in the U.S., meaning over half of all venture dollars were flowing into AI, much of it into generative‑AI infrastructure and applications. (axios.com)
• Barron’s reports that close to one‑third of all VC funding in 2024 went to AI companies, with a heavy concentration in Silicon Valley, explicitly noting AI as the key driver of the modest rebound in VC activity. (barrons.com)

At the same time, mainstream financial and policy commentary now explicitly refers to an “AI bubble” or “AI valuation bubble”: there are warnings about a speculative boom driven by sky‑high valuations, revenue multiples of 20–50x for AI startups, and trillion‑dollar-plus aggregate valuations concentrated in a small set of generative‑AI leaders (OpenAI, Anthropic, xAI, etc.). (ainvest.com) An MIT‑linked study finding 95% of generative‑AI projects failing to show ROI is being cited as classic bubble‑type evidence—massive capital deployment with little near‑term economic return. (timesofindia.indiatimes.com)

Putting this together: since 2023 the dominant Silicon Valley funding narrative and hype cycle has clearly centered on generative AI, with an enormous number of startups formed and widespread concern about a bubble. That matches Friedberg’s forecast about both scale (“100,000 startups”) and the sector becoming the focal point of the next tech hype/bubble cycle, so the prediction is best judged as right.

Friedberg’s prediction has effectively played out so far.

1. Many competitors to Google’s “one‑box” / search-results model have emerged.

   • OpenAI launched ChatGPT Search (originally SearchGPT) in 2024 as an AI search engine explicitly positioned as a direct competitor to Google, Perplexity, and Bing, combining web search with LLM-generated answers and citations. (en.wikipedia.org)
   • Perplexity AI has grown into a well‑funded AI search engine that mixes its own index with LLMs; a 2025 report notes it raising a large new round at a multibillion‑dollar valuation and being cited in the U.S. Google antitrust case as evidence that AI search rivals are real competitors. (barrons.com)
   • Microsoft integrated OpenAI models into Bing and then Copilot, using generative AI features to differentiate Bing search and contributing to noticeable Bing share gains in the U.S. (proceedinnovative.com)
   • Google’s own Gemini, plus other AI chatbots with built‑in web access, have created a crowded field of LLM‑centric interfaces that users can use instead of traditional Google search for many information tasks. (techradar.com)
     Collectively, these confirm the “coming years” have indeed produced multiple serious, LLM-based competitors to Google’s classic results-page paradigm.

2. Google’s core search product is clearly at risk of radical disruption, even if it has not yet been displaced.

   • Google has overhauled Search by rolling out AI Overviews (formerly SGE) worldwide and later AI Mode, which can replace the traditional list of links with comprehensive, Gemini-generated answers. These AI summaries can dominate most of the visible screen and are explicitly framed by Google as a response to generative‑AI competition like ChatGPT. (en.wikipedia.org)
   • The new AI overlays are materially changing traffic flows and business models on the web: companies like Chegg and publishers such as Penske Media have sued Google, alleging that AI Overviews cannibalize their traffic and fundamentally alter search economics. (washingtonpost.com)
   • Industry coverage describes Google as “disrupting itself” to beat OpenAI, reorganizing around Gemini and embedding it directly into Search—behavior consistent with a company that sees its core product as under existential threat from new technology. (businessinsider.com)

3. At the same time, Google’s dominance shows the disruption is still in progress, not complete.

   • As of 2025, Google still processes roughly ~90% of global web searches, far ahead of Bing and all other rivals, and Alphabet’s financial results and market cap indicate that search remains extraordinarily profitable. (mettevo.com)
     This means Google’s search business has not yet been radically overturned in market share or revenue terms, but it is facing credible AI-driven challenges that are forcing a deep redesign of how search works.

Because Friedberg’s claim was that (a) LLM/chat interfaces would generate many competitors to Google’s search-results model, and (b) Google’s core search engine would be at risk of radical disruption—not that it would already be fully disrupted by 2025—current evidence supports his prediction. The competitors and the strategic risk to Google’s core product are real and widely acknowledged, even though Google still dominates usage today.

Chamath predicted that over the coming years many traditional enterprise SaaS apps would be incrementally replaced by “models as a service” (MaaS), leading to a broad shift from SaaS to MaaS.

What we see by late 2025:

1. SaaS is still large and growing, not shrinking.

   • Global SaaS revenue is still rising fast: estimates put the 2024 market around $250–400B, with forecasts to exceed $1T by the early 2030s, and SaaS adoption already above 70% of organizations. These reports describe continued expansion, not a broad replacement of SaaS as a model. (marketdataforecast.com)

2. MaaS / AI-model APIs have clearly emerged as a category.

   • Major cloud vendors now explicitly market “Model(s) as a Service” offerings, providing pre‑trained ML and LLM models via APIs (e.g., Microsoft Azure’s MaaS description, Cassava’s model‑as‑a‑service, Red Hat’s “Models‑as‑a‑Service” in OpenShift AI). (azure.microsoft.com)
   • This matches the direction Chamath described: single‑use/specialized models accessed as cloud services.

3. AI is pressuring traditional SaaS, but evidence of wholesale replacement is isolated.

   • Consulting analysis (AlixPartners) finds over 100 public mid‑market software firms under pressure as generative AI lets lean, AI‑native entrants and big incumbents challenge traditional SaaS models; growth and retention are weakening for some of these companies. (businessinsider.com)
   • Some high‑profile anecdotes (e.g., Klarna reportedly decommissioning ~1,200 SaaS tools, including Salesforce, in favor of an internal AI knowledge platform) show that individual enterprises can replace many SaaS apps with AI‑centric systems. (medium.com)
   • However, these are still case studies, not clear evidence of a broad market‑wide shift from SaaS to MaaS.

4. In practice, AI/MaaS often augments SaaS rather than replacing it.

   • Many SaaS vendors are embedding AI features into their existing products (e.g., HubSpot adding AI‑generated emails, chatbots, predictive analytics) while remaining fundamentally SaaS platforms. (forbes.com)
   • AI model providers (OpenAI, etc.) are mostly used as infrastructure by other apps, though some moves toward direct SaaS competition are emerging; analysis frames this as a new competitive threat, not yet a completed replacement. (businessinsider.com)

Why the verdict is “inconclusive”:

• Chamath’s claim has two parts:

  1. Directional trend: models as a service will become a key way enterprises consume functionality.
  2. Outcome framing: SaaS will be replaced over time, with a broad shift from SaaS to MaaS.

• By late 2025, part (1) looks directionally supported: MaaS/AIaaS is clearly a recognized, fast‑growing pattern and is starting to power many workflows.

• But for part (2), macro data still shows a robustly growing SaaS market, with only early signs of displacement and a few headline examples of companies ripping out large numbers of SaaS tools. There’s no strong, quantitative evidence yet that many traditional enterprise SaaS applications have already been replaced or that SaaS as a whole is in secular decline.

• Because Chamath’s timeframe—“over the coming years” and “over time”—is vague and plausibly extends well beyond 2025, we cannot yet say whether a broad structural shift from SaaS to MaaS will ultimately occur.

Therefore, as of November 2025, the prediction is directionally plausible but not clearly fulfilled or falsified, so the fairest label is "inconclusive (too early)".

By late 2025, the prediction has clearly materialized.

OpenAI released GPT-4V in 2023, adding image understanding to GPT‑4 and explicitly positioning multimodal LLMs (beyond text-only) as a major frontier in AI, with vision tightly integrated into the language model rather than as a bolt-on tool. (openai.com) In May 2024, OpenAI launched GPT‑4o, described as a large multimodal model that can natively take text, images, and audio as input and produce text, image, and audio outputs, aimed at more natural, interactive voice and video experiences. (openai.com) OpenAI has since built real-time voice and agent products (e.g., Realtime API and next‑generation audio models) on top of this architecture, enabling practical, complex applications like voice agents that understand speech, see the screen, and act. (openai.com)

Other big tech companies have independently taken the same path: Google DeepMind’s Gemini family, announced in December 2023, is natively multimodal across text, images, code, audio, and video, and is positioned as Google’s most powerful AI offering, integrated into products like Bard/Gemini apps and planned across Search and Ads. (time.com) The broader ecosystem now routinely deploys vision‑language and multimodal models (e.g., in Microsoft Copilot and similar tools) to tackle substantive, real‑world tasks that go beyond what single‑mode text models could handle. (en.wikipedia.org)

Given that (1) within a few years of the 2022 podcast, major advances did come from OpenAI and other large tech companies, (2) these advances are explicitly multimodal across text, audio/voice, images, and in some cases video, and (3) they are used to solve more complex, real‑world problems, Chamath’s prediction is best classified as right.

Evidence since late 2022 shows that even the best frontier LLMs have not reached the kind of near‑perfect, guaranteeable reliability needed for the last 1–2% of high‑consequence use cases, matching Chamath’s prediction.

1. Vendors explicitly warn against high‑stakes reliance. OpenAI’s GPT‑4 technical report and legal briefings summarizing it state that GPT‑4 “is not fully reliable” and can still hallucinate, and they recommend human review, grounding with additional context, or avoiding high‑stakes uses altogether. These documents stress that care is required “particularly in contexts where reliability is important.” (mondaq.com) OpenAI’s 2025 Operator system card goes further, saying the system proactively refuses high‑risk tasks like banking transactions or other high‑stakes decisions, and enforces a supervised “watch mode” on sensitive sites—clear acknowledgment that current models can’t be trusted to act autonomously in these domains. (openai.com)

2. Persistent hallucinations in critical domains (law, medicine, scientific work).

   • Legal: A 2024 study on “legal hallucinations” finds that when asked specific, verifiable questions about random U.S. federal court cases, ChatGPT‑4 hallucinated in 58% of cases, and other models did even worse, leading the authors to warn against unsupervised legal use. (arxiv.org)
   • Medicine: A 2024 paper on patient‑summary generation reports that even carefully tuned models like GPT‑4 still generate non‑trivial numbers of medical hallucinations, with authors explicitly advising caution for clinical use because standard metrics don’t capture all errors. (arxiv.org)
   • Research summarization: Work on hallucinations in academic paper summaries finds that GPT‑4 and other frontier models regularly insert subtle but incorrect claims; automated methods are required just to detect these, and the authors again recommend caution. (arxiv.org)

3. Hallucinations remain a fundamental, hard‑to‑eliminate problem. Recent coverage and research note that hallucinations are structural to how LLMs work and cannot be fully eliminated with current architectures. A 2024–2025 wave of reporting and studies emphasizes that even as models get more capable, hallucinations persist and can even increase for some new reasoning models (e.g., OpenAI’s o3 and o4‑mini showing higher hallucination rates than an older o1 model). Experts stress that this is especially problematic in domains like law, medicine, and finance where rare errors are unacceptable. (livescience.com)

4. Regulatory and technical consensus that reliability is insufficient for unsupervised high‑risk use. The 2024 EU AI Act explicitly imposes strict robustness, risk‑management, and human‑oversight obligations on “high‑risk” AI, reflecting that current systems are not considered dependable enough for critical applications without strong controls. (en.wikipedia.org) Overviews of LLM limitations continue to list hallucinations and brittleness as key barriers to deployment in high‑stakes settings. (en.wikipedia.org)

By late 2025—nearly three years after the December 2022 podcast—frontier LLMs are dramatically more capable but still cannot reliably deliver the last 1–2% of precision required for autonomous use in high‑consequence domains, and major developers and regulators openly treat this as an unsolved, exceptionally hard problem. That aligns closely with Chamath’s prediction, so it is right for the (now‑elapsed) near‑to‑medium‑term window he was talking about.

Chamath makes a time‑to‑solve prediction: that going from ~98–99% to near‑perfect reliability in complex AI systems (like self‑driving or high‑stakes inference) will take “decades”. In the episode transcript he explicitly says these “last 100 or 200 basis points literally takes decades,” in the context of self‑driving and brittle AI models. (podscripts.co)

As of November 2025:

• Fully general SAE Level 5 self‑driving (any road, any conditions, no human fallback) still has not been achieved or widely deployed. Major players like Waymo and Cruise remain at limited geofenced robotaxi services with notable safety and regulatory problems (e.g., Cruise’s California permits suspended and its driverless operations paused after a serious pedestrian‑dragging incident, then later GM effectively exiting the robotaxi business). (cnbc.com)
• At the same time, systems have improved sharply: Waymo reports tens of millions of driverless miles with substantially fewer insurance claims and injury crashes than human drivers, showing rapid progress toward that high‑reliability regime rather than stagnation. (nbcbayarea.com)

However, none of this tells us how long it will ultimately take to close the final 1–2 percentage points of reliability. Only about three years have elapsed since the prediction in December 2022; “decades” implies a horizon on the order of 20+ years. We cannot yet know whether the remaining progress will in fact stretch over multiple decades or be compressed into a shorter period.

Because the forecast is specifically about a long future time scale and we are still near the beginning of that period, the correctness of the claim cannot yet be determined.

So the appropriate classification is: it’s too early to tell → inconclusive.

Evidence from 2023 shows that both elements of Sacks’ prediction broadly played out over the following year.

1. “Big contraction in jobs … around this industry”
   Tech layoffs in 2023 were very large, and a significant share were at cloud and SaaS firms. Layoffs.fyi–based tallies cited by multiple outlets show over 160k–220k tech layoffs in 2023, exceeding 2022’s totals.(indigilife.com) Major enterprise SaaS vendors—including Salesforce (10% of workforce), Atlassian (5%), Twilio (~17%), Zoom (15%), and others—are specifically listed in 2023 layoff roundups.(joltjournal.com) SaaS-focused reporting likewise notes that many SaaS firms had rapidly over-hired 2–3x in prior years and then moved to layoffs and hiring freezes as conditions worsened.(moneycontrol.com) This supports the claim that there was a big contraction in jobs in and around the SaaS sector during the year after December 2022.

2. “Vicious cycle … where seat contraction becomes the norm instead of seat expansion”
   On the customer/seat side, multiple data points and company disclosures show a shift from the prior decade’s automatic seat expansion to flat or negative seat counts at many SaaS vendors:

   • A March 2023 TechCrunch analysis of $2.5B of SaaS spending across 18,000 deals reported that recent layoffs had caused a “decline in seat licenses”, and concluded that in 2023 “a flat renewal is the new ‘upsell’” and that SaaS vendors should expect contraction at renewal, not expansion, as the immediate impact of layoffs on seat counts.(techcrunch.com) This is essentially the dynamic Sacks described.
   • Expensify’s Q4 2023 results broke out paid-seat movement: in 2022, existing customers added ~85,000 seats, but in 2023 the same cohort lost ~42,000 seats, with management calling 2023 “a brutal year” for customers and attributing the swing to customers hiring then laying off staff.(marketbeat.com) That is a textbook case of seat expansion turning into net seat contraction within a year.
   • A SaaStr breakdown of Dropbox at ~$2.5B ARR noted that seat contraction and churn remained elevated, with customers reducing seat counts even as the company pushed modest ARPU increases.(saastr.com) Again, this aligns with seat contraction/rightsizing at renewal.
   • The 2023 Zylo SaaS Management Index found that organizations on average waste or underutilize ~44% of their SaaS licenses, with large enterprises only using about half of their purchased seats. It frames cutting redundant and unused SaaS licenses as a primary lever to reduce costs during 2023’s wave of layoffs, reinforcing that many customers were actively shrinking or optimizing seat counts rather than expanding.(prnewswire.com)

   At the same time, industry-wide SaaS spending and revenue still grew in 2023—Gartner and SaaS-market analyses projected double‑digit increases in software spending, and one large dataset expected SaaS spend to grow about 18% that year.(techcrunch.com) So the sector was not in an outright collapse. But Sacks did not predict a revenue death spiral; he specifically focused on:

   • headcount-driven job cuts in and around SaaS, and
   • the shift from the long‑standing baseline of automatic seat expansion to seat contraction/flat renewals over “the next year or so.”

Given the scale of 2023 tech/SaaS layoffs and the documented move toward flat or negative seat counts and downsized renewals at many SaaS vendors—while acknowledging that effects varied by company—the prediction is directionally well‑supported by the data. Hence the verdict: right.

Industry data for 2023 show that, while SaaS expansion slowed and churn/downsells increased, the baseline did not shift to 80–90% of prior-year revenue from existing customers.

• SaaS Capital’s 2023 retention benchmarks (one of the largest private B2B SaaS surveys) report median net revenue retention (NRR) of 102% and median gross retention of 91%, unchanged from 2022. That implies the typical SaaS company still grew revenue from its existing customer base rather than shrinking it. (saas-capital.com)
• A summary of the same data notes the median NRR across all SaaS companies was 102%, with higher ACV enterprises at around 110% NRR—again consistent with modest expansion, not 10–20% contraction. (churnzero.com)
• KeyBanc/Sapphire’s multi-year private SaaS survey finds that net retention stayed above 100% through the downturn, even though it fell from 2021 peaks; what deteriorated more sharply was gross retention, which dropped to the mid‑80s in 2023 before starting to recover. (prnewswire.com)

Some weaker or SMB-focused SaaS companies did experience NRR below 100% and heavy seat reductions, but broad-based surveys show:

• Typical/median SaaS did not fall to 80–90% of prior-year revenue from existing customers.
• Net retention in 2023 remained slightly above 100% for the sector as a whole.

Because the prediction framed this level of contraction as the new baseline for “many” SaaS companies, contrasting it with the historical 120%+ NRR, it overshot the actual severity of the downturn and is best characterized as wrong.

This prediction is qualitative and relative, so it can’t be cleanly labeled right or wrong.

Evidence on the 2022–2023 SaaS environment:

• Public SaaS valuation multiples and growth expectations compressed sharply after late 2021. Bessemer’s Cloud Index data shows revenue multiples and growth rates for public cloud companies falling through 2022–2023 compared to the 2020–2021 boom, reflecting a much tougher funding and demand environment for SaaS overall.
• Venture funding for SaaS and broader startup markets dropped substantially in 2022–2023 relative to 2020–2021, with many reports describing a “reset” and investors prioritizing efficient growth over growth-at-all-costs. This supports the idea that sustaining very high growth (e.g., 2x year-over-year) became notably harder and more prized versus the prior frothy period.
• However, some SaaS companies (especially AI-adjacent or infrastructure players) continued to post extremely high growth rates even in 2023–2024, sometimes exceeding 2–3x year‑over‑year from smaller bases. At the same time, other later‑stage SaaS companies struggled to grow at all. This dispersion makes it impossible to define a single, universally accepted conversion like “2x now ≈ 3x then” across the entire category.

Reasoning:

• The directional claim—that it became significantly harder and more impressive to hit 2x growth in 2022–2023 than to hit 3x in the prior boom—is broadly consistent with macro data on SaaS markets: lower valuations, tighter funding, and slower average growth.
• But the specific equivalence ("2x now is as good as 3x then") is a normative, relative judgment that depends on:
  • which segment of SaaS (early‑stage vs late‑stage, infra vs application),
  • which investor or operator you ask, and
  • what baselines they use for "impressive" in each period.
• There is no objective benchmark or consensus metric that could definitively verify or falsify this exact conversion ratio.

Because of this, even with enough time having passed, the prediction can’t be conclusively validated or refuted from available data; it’s a qualitative calibration of difficulty and impressiveness rather than a clearly testable forecast.

Therefore the appropriate classification is: "ambiguous".

By late 2025, industry consensus and market behavior largely match Chamath’s thesis. Major vendors now describe LLMs as heading toward commodity status, with Microsoft’s AI leadership explicitly warning that as models rapidly converge in capability, they will be quickly commoditized and that the real value for businesses lies in how models are integrated with their own data and workflows. (microsoft.com) Analysts likewise point to steep token-price declines and argue there is little durable moat in the base models themselves, characterizing the space as undergoing rapid commoditization. (linkedin.com)

This is reinforced by the proliferation of strong open- and closed-source foundation models (e.g., Llama, Mistral, DeepSeek) that are either open-weight or easily accessible via APIs, making high-quality general-purpose capabilities widely available rather than exclusive to a single firm. (blog.gordonbuchan.com) Research on Llama 2’s release even treats it as a “commoditization shock” that lowered barriers for downstream developers, consistent with his view that general-purpose models would become baseline infrastructure. (aisel.aisnet.org)

At the same time, business and strategy literature now repeatedly frames proprietary data as the key competitive moat in generative AI—calling it “the new gold,” “the only sustainable enterprise AI moat,” and “the decisive factor” separating winners from laggards. (forbes.com) IBM and California Management Review similarly stress that, because large foundation models are broadly available, advantage comes from customizing them with enterprise-specific data and embedding them in differentiated workflows. (ibm.com) Some executives still argue that frontier models are not yet fully commoditized, but even those perspectives focus on differentiation via systems and domain-specific data rather than the generic base models. (linkedin.com) Overall, the observed direction of the market—models trending toward interchangeable building blocks, with competitive edge shifting to proprietary data—aligns closely with Chamath’s prediction, so it is best judged as right in substance, even if commoditization remains an ongoing process rather than fully complete.

Evidence since (and even before) late 2022 shows that AI models for tumor detection trained on large proprietary image corpora have indeed been cleared by the FDA relatively quickly via existing pathways.

Examples in breast cancer imaging

• Lunit INSIGHT MMG, an AI system for breast cancer detection, was trained on over 240,000 mammography cases including up to 50,000 with breast cancer and received FDA 510(k) clearance in November 2021—before Chamath’s comment. This is exactly the kind of large‑corpus, tumor‑classification AI he described, cleared via the standard 510(k) pathway, not a special new regime. (lunit.io)
• Therapixel’s MammoScreen, an AI-based breast cancer screening aid, received 510(k) clearance in 2020 and a second 510(k) in 2021 extending it to 3D tomosynthesis, again after multi‑reader clinical studies but through normal device pathways. (therapixel.com)
• DeepHealth/RadNet’s mammography AI products (Saige‑Q and later SmartMammo/SmartMammo Dx) also obtained 510(k) clearances, first in 2012 and then expanded indications and system compatibility in 2022 and 2024, illustrating repeated use of the same regulatory route for breast imaging AI. (diagnosticimaging.com)

Use of existing FDA pathways and speed

• Systematic reviews of FDA AI/ML-enabled devices show that by October 2023 there were ~691 such devices authorized, with about three‑quarters in radiology and the vast majority cleared via the traditional 510(k) pathway based on substantial equivalence rather than bespoke rules. (mdpi.com)
• By August 2024, FDA’s own updates report 950 AI/ML-enabled medical devices cleared or approved, again emphasizing standard pathways (510(k), De Novo, PMA) for AI tools, especially in imaging. (fda.gov)
• A study of AI/ML-enabled surgical devices found average FDA review times of roughly 142 days (~4.7 months), in line with typical 510(k) timeframes and supporting the characterization of these pathways as relatively fast compared with full Premarket Approval processes. (asc-abstracts.org)

Assessment vs. the prediction Chamath’s claim had two parts:

1. Such AI models could be approved using existing regulatory pathways. This is clearly borne out: AI mammography/tumor‑detection systems have repeatedly obtained 510(k) clearances without the FDA needing to create a fundamentally new pathway.
2. Those pathways allow approval “very quickly.” While “very quickly” is subjective, the empirical review times of a few months for 510(k) AI devices, versus years for full PMA processes, plus the rapid growth to hundreds of cleared radiology AI tools, support that the process is relatively fast in regulatory terms.

Because multiple AI systems closely matching his example (breast cancer image classifiers trained on very large datasets) have gained FDA clearance through existing, comparatively rapid pathways, the prediction that this kind of AI would be able to get FDA approval relatively quickly via existing mechanisms is best judged as right.

Chamath’s normalized prediction is that in the near term (i.e., the few years after late 2022) the largest purchasers of machine‑learning hardware would be the U.S. government and ultra‑high‑frequency trading (HFT) firms.

Evidence from 2023–2025 instead shows that the dominant buyers of AI/ML hardware (especially GPUs/accelerators) are hyperscale cloud and Big Tech companies, not governments or HFTs:

• Multiple analyses of Nvidia’s customer base identify Amazon, Microsoft, Alphabet/Google, Meta, and Oracle as the biggest buyers of Nvidia’s data‑center AI GPUs. The U.S. government and HFT firms do not appear on these lists. (fool.com)
• Research and reporting on Nvidia’s H100/Hopper boom show that Microsoft and Meta were among the very largest buyers, each acquiring on the order of 150,000 H100 GPUs in 2023 alone, with Amazon and Google also buying tens of thousands. These four Big Tech companies together account for roughly 40% of Nvidia’s revenue, indicating that hyperscalers, not governments or HFTs, dominate demand. (observer.com)
• Capital‑expenditure disclosures and analyst estimates show tens of billions of dollars per year in AI infrastructure spending from Microsoft, Amazon, Google, and Meta—e.g., Microsoft and Amazon each spending on the order of $30–80B+ in annual capex with a large fraction dedicated to AI chips, servers, and data centers. (fool.com)

By contrast, while U.S. government AI spending is rising quickly, it is much smaller and not concentrated purely in hardware:

• A Brookings‑cited analysis found U.S. federal AI‑related contract values grew from about $355M to $4.6B between August 2022 and August 2023, driven mostly by the Department of Defense. Yet the same reporting notes that private‑sector AI investment vastly outstrips public funding, highlighting companies like Meta and Microsoft spending billions annually on AI infrastructure such as high‑performance GPUs. (time.com)
• Large cloud contracts for U.S. government workloads (e.g., AWS’s newly announced $50B AI/HPC build‑out for federal customers) show that even government demand is primarily intermediated through hyperscalers—AWS, Azure, etc.—who themselves are the direct massive purchasers of the hardware. (reuters.com)

For ultra‑high‑frequency/quant trading firms, there are examples of significant GPU clusters—e.g., China’s High‑Flyer hedge fund built a system with about 10,000 Nvidia A100 GPUs—but this is still tiny compared with the hundreds of thousands to millions of GPUs being accumulated by a single hyperscaler like Microsoft. (en.wikipedia.org) There is no credible market data suggesting that HFT firms, in aggregate, rival the big cloud/AI players in total machine‑learning hardware purchases.

Given that nearly three years have passed since the Dec 2022 podcast and the clear consensus of both industry reporting and financial data is that hyperscalers/Big Tech—not the U.S. government or HFTs—are the largest buyers of ML hardware, Chamath’s normalized prediction about who would remain the biggest buyers in the near term is wrong.