E168: Can Google save itself? AI takes over Customer Support, Reddit IPO teardown

Available data show that the condition Chamath described has effectively occurred, but the market reaction he predicted did not.

1. Google’s search share has fallen from the low‑90s to the high‑80s.

   • StatCounter-based summaries put Google’s global search share at about 91.5% in January 2024.【1†turn1search10】
   • By late 2024, StatCounter data showed Google dropping below 90% for three consecutive months, hitting 89.34% (Oct), 89.99% (Nov), 89.73% (Dec).【1†turn1search5】
   • Business Insider, citing StatCounter, reports Google’s global share fell from ~93% in March 2023 to 89.71% in March 2025, and from about 91% in March 2024 to 89.71% in March 2025.【4†turn4view0】 Statista likewise puts Google at 89.62% in March 2025.【1†turn1search3】
   • Other 2024–25 overviews describe Google’s share as ~89–90% worldwide, confirming a sustained dip of a few percentage points from the earlier 91–93% range.【0†turn0search7】【0†turn0search9】

   Taken together, this is roughly the 3‑percentage‑point erosion (from about 92–93% down to ~89–90%) that his scenario (“300–500 basis points of change”) contemplated.

2. Alphabet’s market cap did not get cut in half; it surged to record highs.

   • Alphabet’s end‑of‑year market cap was about $1.76T in 2023, $2.36T in 2024, and is estimated around $3.87T for 2025, per compiled market‑cap histories.【2†turn2search0】【2†turn2search4】
   • Real‑time and news coverage in late 2025 describe Alphabet approaching a $4T valuation, with the stock up roughly 70% year‑to‑date and near all‑time highs.【0†turn0news12】【2†turn2news12】

Given that (a) Google has indeed experienced on the order of a few‑percentage‑point drop in global search share from its prior ~92–93% dominance, yet (b) Alphabet’s market capitalization has risen dramatically instead of being cut in half, the specific prediction that a 3–5 point share loss would trigger a ~50% market‑cap collapse is not borne out by actual market behavior so far.

Because the antecedent (a multi‑point share decline from the low‑90s) has effectively occurred and the predicted consequence (halving of market cap in short order) clearly has not, the forecast is best classified as wrong, rather than merely “too early to tell.”

Chamath’s claim was that if Perplexity or any other search engine "clips off" about 50–100 basis points (0.5–1.0 percentage points) of search-market share from Google, Alphabet’s stock would "go straight down by 50%" once that loss shows up in the data.

What happened to Google’s search share?

• StatCounter-style data show Google’s worldwide search share (all devices) falling from about 91.6% in February 2024 to roughly 89.6–90.4% in 2025, i.e., a decline of about 1–2 percentage points. Over the same period, Bing’s share rose from about 3.3% to ~4.0%, with smaller gains for Yahoo and DuckDuckGo. (proceedinnovative.com) That is exactly the kind of 50–100 bps competitive gain Chamath described.
• However, the specific AI upstart he named, Perplexity, still does not register anywhere near 0.5–1.0% of global search-engine share; industry writeups note that Perplexity’s share of general web search is below major dataset detection thresholds (<0.1%), even if it has a few percent of the much smaller "AI search" niche. (vedranmarkovic.com)

What happened to Alphabet’s stock?

• On the podcast date (1 March 2024), Alphabet Class A (GOOGL) closed at about $136 per share. (statmuse.com) By late November 2025 it trades around $320 and is near all‑time highs, with Alphabet’s 2025 year‑high closing price above $320. (macrotrends.net)
• Over 2024–2025, the worst drawdowns visible in daily/annual data are on the order of 20–30% from recent highs, not anything close to a 50% collapse. (macrotrends.net) In fact, instead of crashing, Alphabet’s market capitalization rose to about $3 trillion for the first time in September 2025, at then‑record stock prices. (investopedia.com)

Assessment Even after Google visibly lost more than the 50–100 bps of search share Chamath highlighted (mostly to Bing and other incumbents), the stock did not "go straight down by 50%"; it roughly doubled from the podcast date and set new highs. Because the antecedent (a ~0.5–1.0 percentage‑point competitive share gain at Google’s expense) has effectively occurred while the predicted 50% price crash clearly has not, this prediction is best scored as wrong. The only caveat is that Perplexity itself hasn’t yet taken that much global share—but Chamath explicitly said "Perplexity or anybody else," and that broader version has already been falsified by the data.

Evidence by late 2025 shows that licensing proprietary datasets to AI model providers has become a standard, incremental revenue stream for many operators of websites and apps with unique data, matching Chamath’s prediction.

• Reddit explicitly frames AI data licensing as a new business line: its IPO filings describe aggregate contracts of about $203M for licensing user-generated content to AI firms (including Google and later OpenAI), with at least $66.4M recognized as 2024 revenue. Coverage emphasizes that these data deals are a distinct, fast‑growing revenue stream alongside advertising, not a one‑off windfall. (techcrunch.com)
• News publishers and media platforms (e.g., News Corp, Le Monde, Dotdash Meredith, Future, Axel Springer, Financial Times, Prisa Media, The Atlantic) have all signed multi‑year content licensing agreements with OpenAI and others, with reporting explicitly describing these as recurring revenue or “new revenue streams” built on their proprietary archives. The News Corp–OpenAI deal alone is reported to be worth over $250M across five years. (apnews.com)
• Broader cross‑industry trend: Surveys of 2024–2025 “proprietary data licensing” deals show a growing market where many data‑rich operators monetize specialized datasets—Reddit (social discussion), Shutterstock (images), Google–Stack Overflow (programming Q&A), Apple–Shutterstock (images), Tempus (clinical/genomic data), Meta–Reuters (news), and others—specifically for AI training and model improvement. These are presented as structured, repeatable commercial arrangements, not experimental pilots. (datafaire.ai)
• Smaller operators and entrepreneurs are participating. Market reports and Reuters coverage describe companies like Defined.ai aggregating niche, high‑signal datasets (e.g., medical imagery, crime‑scene photos, specialized audio) from individual entrepreneurs and content owners, then licensing them to major AI developers. The revenue is shared with those data owners, demonstrating that even relatively small operators with unique data can now sell into AI‑training pipelines. (investing.com)
• Regulatory and legal disputes reinforce that licensing is now the expected path. Lawsuits by Reddit against Anthropic and Perplexity explicitly contrast alleged unauthorized scraping with Reddit’s licensed deals with Google and OpenAI, underlining that selling access to the dataset is a normal, monetized option that counterparties are expected to pay for. (investopedia.com)

Taken together, by November 30, 2025 (less than two years after the March 2024 episode), there is a broad and growing ecosystem where operators of sites and apps with unique, high‑quality data routinely license that data to AI model providers as a distinct revenue line. While not every entrepreneur will succeed in doing so, it is clearly and commonly possible—and already happening at scale—so the core economic prediction is best judged as right.

As of November 30, 2025, the time horizon of the prediction (“within the next several years”) has not elapsed, so it’s too early to judge its long‑term accuracy.

Reddit: Reddit went public on March 21, 2024, and continues to operate as an independent public company (ticker RDDT). Recent coverage and filings describe it as a standalone public firm with a diversified institutional shareholder base, ongoing earnings reports, and strategic partnerships, not as a subsidiary that has been acquired in a take‑private or merger transaction. (en.wikipedia.org) There is no evidence of Reddit being bought by a larger company after its IPO.

Stack Overflow: Stack Overflow (via Stack Exchange) has been wholly owned by Prosus—a large tech holding company—since its acquisition in 2021, well before the March 2024 podcast. (en.wikipedia.org) That prior deal doesn’t test the future part of the prediction.

Because (1) Reddit has not been acquired yet, and (2) the multi‑year window the speaker invoked is still open, we cannot say the prediction is definitively right or wrong at this point.

Within roughly a year of March 1, 2024, there was a clear shift from mostly unpriced, implicit scraping of web content toward a recognizable market for training data, with repeated, high-profile licensing deals that put concrete dollar values and contractual patterns around content used for large language models.

From spring 2024 onward, OpenAI and other AI firms signed a series of explicit, paid content-licensing agreements with major news organizations and platforms: Le Monde and Spain’s Prisa Media in March 2024; the Financial Times in April 2024; and a large, multi‑year News Corp deal in May 2024 that press reports said could exceed $250 million over five years, following earlier agreements with the Associated Press and Axel Springer. These contracts typically cover both training use and in-product display, and industry reporting indicates that OpenAI and peers now routinely pay anywhere from low single‑digit millions per year for smaller publishers up to nine‑figure packages for marquee groups, establishing de facto price bands for premium text data. (openai.com) In parallel, platform data deals such as Reddit’s reported $60 million‑per‑year agreement with Google for API access, plus subsequent AI partnerships, reinforced the notion that large conversational datasets have substantial, quantifiable market value. (time.com) Broader industry analysis by mid‑ to late‑2024 described a “data gold rush” in which big tech firms were systematically hedging legal and supply risk by buying or licensing training data (e.g., Shutterstock image/video/music libraries, Defined.ai brokered datasets) at increasingly standardised per‑asset or per‑word prices, further clarifying commercial norms for AI training data. (wifc.com)

At the same time, the legal framework around copyright and fair use for training remained unresolved, with major lawsuits such as The New York Times and other publishers vs. OpenAI and Microsoft allowed to proceed into discovery in 2024–2025 rather than being definitively settled, and new legislation like the U.S. Generative AI Copyright Disclosure Act only beginning to address transparency obligations. (apnews.com) That ongoing litigation shows the law is not fully settled, but it does not negate the prediction: by early 2025, there was a much clearer, widely reported pattern of how AI companies and rights‑holders were valuing and transacting over training data (license-or-litigate, with well‑understood deal structures and price ranges), compared with the far murkier situation in early 2023–early 2024. Given that the claim was that things would start to become clearer over the following year, not that every legal and economic question would be fully resolved, the prediction is best judged as having come true.

• What was predicted. Chamath said that if two or three more large companies publicly shared Klarna‑style AI customer‑support results with clear, quantitative gains, then Teleperformance (a large call‑center/BPO firm) would become “a $1 billion company in short order.”

• Did the precondition happen? Since Klarna’s early‑2024 disclosure that its AI assistant handled about two‑thirds of customer‑service chats and replaced work equivalent to ~700 agents with faster resolution times and fewer repeat inquiries, multiple other large companies have publicized similarly substantial AI‑driven customer‑support efficiency gains with concrete metrics. (reddit.com)
– H&M has reported that its generative‑AI chatbot cut customer‑service response times by up to ~70%. (aiprm.com)
– Bank of America has highlighted that its virtual assistant Erica has handled around three billion customer interactions and taken on work equivalent to roughly 11,000 employees, explicitly framing this as a major AI productivity lever in customer service and banking operations. (reuters.com)
– Lyft disclosed that its Claude‑powered AI customer assistant cut average support resolution time by about 87%. (datagrom.com)
– Other well‑known firms (e.g., Octopus Energy, Vodafone, Sephora, Bank of America again via Erica) are cited in industry roundups as having materially reduced customer‑service workloads or response times using AI assistants. (linkedin.com)
These meet the spirit of “big companies” publicly touting measurable AI customer‑support productivity gains, so the condition of the prediction is effectively satisfied.

• What actually happened to Teleperformance’s market cap.
– Around the time of the prediction (late 2023), Teleperformance’s market cap was roughly €8.0–8.1 billion (≈$8.9B) at year‑end 2023. (stockanalysis.com)
– By year‑end 2024 it had fallen to about €4.9 billion (≈$5.1B), and by November 28, 2025 it was about €3.4 billion (≈$3.9–4.0B). (stockanalysis.com)
– Historical data from multiple sources show no point in 2024–2025 where Teleperformance’s market cap approached $1 billion; instead it has remained several times higher, even after a substantial multi‑year decline.

• Assessment.
– The trigger Chamath described (multiple big companies broadcasting Klarna‑like AI customer‑support productivity wins with hard numbers) has, in broad economic and public‑disclosure terms, occurred.
– The outcome he forecast—Teleperformance rapidly collapsing to roughly a $1B market cap once that happened—did not occur; as of late 2025 Teleperformance is still around $3.5–4.0B. (stockanalysis.com)

Because the condition has been met while the predicted consequence has clearly failed to materialize even well beyond any plausible “short order” window, this prediction is best judged as wrong.

Jason’s scenario implied that if AI customer-support systems improved quickly enough, they could actually be handling ~98% of customer queries by the end of 2024. In reality, available data from 2024–2025 shows automation and AI assistance rising fast, but nowhere near 98% of queries being fully handled by AI.

For example, Comm100’s 2025 Live Chat Benchmark (based on 220M+ interactions) reports that AI chatbots handled about 73.8% of live-chat interactions in 2024, not 98%. More importantly, only about 45–46% of chats were fully resolved by chatbots alone in both 2023 and 2024, meaning most conversations still required a human agent to finish the job. (comm100.com)

Industry-wide summaries likewise talk about 60–85% of customer interactions being handled or assisted by AI/automation by 2024, and projections of ~75% handled by AI by 2025—again, well below 98% complete automation. (gitnux.org) Even leading-edge deployments aren’t there: Reuters reports India-based LimeChat’s bots currently handle about 70% of customer complaints for clients and aim for 90–95% within a year, i.e., still short of 98% and only for particular customers, not the whole market. (reuters.com)

User-experience studies also show that many customers still feel they need humans to get issues properly resolved: a 2024 UK survey found 54% of users say their problems are only truly solved by speaking with a human, and 40% would pay extra to avoid dealing with AI chatbots. (techradar.com) This is inconsistent with a world where ~98–99% of queries are fully resolved by AI.

The second part of Jason’s statement (reaching 98–99% in about four years if improvement is only ~10% per year, i.e., around 2028) concerns a future date and can’t yet be directly tested. But the concrete, time-bound claim that AI would be doing about 98% of queries by the end of 2024 clearly did not happen. Therefore, overall this prediction is best classified as wrong.

Chamath’s claim is about feasibility (“a one person company should be able to…”), not that most or all founders will do this. That can be evaluated as soon as clear examples exist.

1. Documented solo SaaS founders are already reaching product–market fit with minimal capital.

   • ScreenshotOne, a screenshot‑API SaaS, is run by a 100% bootstrapped solo founder (Dmytro Krasun) and has grown to roughly $12k–$20k MRR and hundreds of customers with no external funding, which is widely taken as product–market fit for a niche B2B SaaS. (reddit.com)
   • A solo founder bootstrapped a feedback‑widget SaaS from a weekend build to $8,200 MRR in 14 months, with ongoing expenses of about $1,400/month (hosting, Stripe fees, a VA). Over that timeline, total out‑of‑pocket spend is far below “a few hundred grand,” yet the product reached thousands in recurring revenue and a $285k exit, indicating strong market validation. (reddit.com)

2. AI, low-cost infrastructure, and tooling have made one‑person companies building real SaaS products routine rather than exceptional.

   • A 2025 piece on “The Solo AI Founder” notes that in 2024 about 36% of founders are solo, and emphasizes that accessible AI tools let a single developer build and launch an AI SaaS product in months. It highlights SiteGPT, built by one developer over a weekend and grown to around $15k MRR as a one‑person SaaS. (plainenglish.io)
   • A 2025 analysis of the “micro‑SaaS revolution” describes an explosion of one‑person SaaS businesses, explicitly crediting AI and automation for enabling a single person to run an entire SaaS company, and profiles solo founders like those behind Base44 and Writesonic who bootstrapped to large ARR without heavy external capital. (msthgn.com)

3. These examples existed by 2024–2025, well within the “next few years” window. By March 2024 (when the prediction was made) such solo, low‑capital paths to product–market fit were already being achieved and have only become more common by late 2025, with many credible, public case studies of solo, bootstrapped founders hitting meaningful MRR and PMF‑like traction on well under a few hundred thousand dollars of total spend. (sidetool.co)

Because the prediction only requires that this path be realistically possible for a one‑person company, and multiple real‑world cases now demonstrate exactly that, the forecast is best scored as right.

Evidence since mid‑2024 shows a clear wave of AI adoption specifically in phone-based customer support, matching Sacks’ prediction about where AI in support would “go next.”

• Numerous vendors now market AI phone agents / voicebots explicitly as IVR and call‑center replacements, offering natural, multi‑turn conversations over the phone (e.g., Air AI, Insighto.ai, Five9 IVA, Talkdesk AI, Callin.io, Phonecall.bot). These platforms emphasize handling inbound and outbound calls, appointment scheduling, order support, and triage—functions traditionally done by IVR trees and human call‑center agents. (siit.co)
• Industry and vendor blogs in 2024–2025 describe AI voice agents as a major new wave in contact centers: replacing rigid IVR menus with conversational agents and positioning AI receptionists/phone agents as a replacement for traditional call centers, especially for SMBs. Articles talk about AI voice receptionists “replacing traditional call centers” and AI phone calls becoming the “new normal” in 2025, with SMBs and enterprises deploying these agents at scale. (blog.peakflo.co)
• Contact‑center commentary notes that voice remains over half of all customer interactions, and AI is increasingly applied specifically to voice to make calls feel like natural conversations instead of menu trees—exactly the IVR‑to‑AI transition Sacks described. (techradar.com)
• On realism and whether callers realize they’re speaking to AI: AI‑phone vendors highlight hyper‑realistic, human‑sounding voices as a key feature, and at least one high‑profile case (Phonely’s AI agents, deployed in real contact‑center environments and used to replace hundreds of human agents) reports accuracy above 99% and explicitly claims that customers “can’t tell they’re not human.” (venturebeat.com)
• Research and product releases focused on low‑latency LLM‑based voice agents for telecom and call centers further confirm that significant R&D and deployment are now aimed squarely at real‑time phone support, not just text chat. (arxiv.org)

Adoption is not universal and many centers still use humans or hybrid human+AI. But by late 2025, there is a clear, industry‑recognized wave of AI phone agents replacing or front‑lining IVR/call‑center interactions, and several deployments show voices realistic enough that many callers do not realize they are speaking to AI. Directionally and qualitatively, Sacks’ prediction has come true.