Why Apple Needs a Management Overhaul & Why Google is Catching Up with Hyperscalers
Summary
Benchmark’s latest partner loss says more about the leverage of elite individual investors than about Benchmark’s durability. Harry called it stunning that someone can hold “one of the top gigs in venture” and still decide it is not enough; Jason warned that Benchmark’s brand remains “freaking powerful,” while Harry argued Victor’s relationships, multistage appetite and ability to retain 100% of the carry make independence rational.
LPs are financing exceptions to long-standing rules on small funds, team cohesion, board service and stage discipline when an exceptional investor asks them to. Jason’s hedge matters: some discarded rules are obsolete, but others may reveal their wisdom in the next downturn. Elad Gil embodies the override: access to category leaders, the ability to pick from pre-seed through late stage and a 10-to-15-year record let him effectively say, “Thank you for your advice. Now, thank you for your money.”
Fast deployment creates immediate relevance, but only correct deployment sustains it. More checks bring more deal flow and markups; bad selection merely produces “relevance in the short term and failure in the long term,” as Tiger and SoftBank demonstrated. The more durable model combines elite early-stage access with huge follow-ons—the playbook Rory sees at Thrive, Greenoaks, Founders Fund and Elad Gil.
Anthropic’s valuation reset from $100 billion toward $150 billion–$180 billion reflects apparent revenue reacceleration and extraordinary developer demand. The discussed trajectory was roughly $1 billion late last year to $4 billion this year, leaving a “brutal” two-horse race with OpenAI. Jason’s central call is that top developers could consume $8,000–$10,000 monthly—not $200—because agents can work in parallel around the clock.
Google is the panel’s best-executing hyperscaler, while Microsoft’s loss of developer-tool leadership is the clearest incumbent failure. Jason’s provisional trade was buy Google and short Amazon: Google Cloud is smaller but fastest-growing, while AWS appears the hyperscaler laggard. Microsoft began with GitHub Copilot yet let Cursor reach roughly $900 million against Copilot’s cited $500 million.
There is still no visible break in AI infrastructure demand, but capex is running much faster than monetization and eventually must clear the income statement. The panel contrasted roughly $600 billion of investment with a $30 billion growing revenue line and noted that the chips will eventually be depreciated; Jason conceded that shorting the thesis early would have taken Nvidia from $100 to $170 against you. “The market can stay irrational longer than you can stay solvent.”
The trillion-dollar incumbents are rich, but the panel sees them as being on the back foot rather than too powerful in AI. Google has a working model; Apple “doesn’t even have a product that works”; Microsoft bought access it does not fully own; and Meta is spending from a “terrible psychological need for a product.” Apple’s hardware and App Store toll may protect it, but Harry—despite Apple being his largest position—said five years without growth should force a board-level management question.
The closing forecasts favored Cursor above $4 billion and Lovable above $400 million by end-2026, while OpenAI at $800 billion split the panel. Harry leaned under; Jason argued over on GPT-5, Codex and a first Jony Ive product. Jason also argued that if OpenAI needs $800 billion, deeply committed investors and sovereign capital will “solve for it.” AI capex already equals about 1.2% of GDP, above the cited 1.1% bandwidth peak but far below the railroad boom’s 6%—enough room for further spending, but also for a painful reset.
Deep dive
1. Benchmark’s churn validates the solo investor’s new leverage
Jason’s first principle was that venture should not expect stability while elite AI researchers jump between Meta, Anthropic and others for enormous packages. For an investor with a hot hand, leaving early can be economically rational: waiting rarely preserves much carry, while the departing partner still has to start over.
Harry separated Benchmark’s prospects from the industry signal. The firm has survived for 30 years, retains a compelling equal-carry, hierarchy-free structure and can recruit replacements; the remarkable fact is that someone can spend two years in “one of the top gigs in venture” and conclude, “No, it’s not enough. I need something more.”
Harry’s evidence for institutional resilience was Benchmark’s recent AI portfolio: Mccor above $100 million, HeyGen around $100 million, Fireworks around $140 million, Sierra with Brett Taylor above $100 million, plus Manus AI and Lorra, all described as double-digit ownership positions in one fund. “How could I be an LP in that fund?”
Victor’s strategic mismatch with Benchmark also mattered. Harry described him as a multistage investor who wants to meet founders from Brex at pre-seed through a $55 million HeyGen check—latitude a disciplined Benchmark model may resist—with the prospect of retaining 100% of the resulting carry.
2. Brand gets a solo fund started; scale of capital changes what it can become
Jason’s caution came from experience: despite already having the SaaStr and 20VC brands and receiving approaches from top firms, he later appreciated how much credibility an institutional name supplies in founder meetings. “Understand the value of the brand you’re leaving behind”; without an individual franchise above the line, independence is difficult.
Harry agreed on the general rule but placed Victor inside an unusually powerful Silicon Valley network alongside Sarah Guo and Elad Gil. That “micro-brand of relationships” can replace institutional signaling, attract private capital and preserve deal flow through the transition.
Once individual and firm brands are functionally equivalent, Harry sees capital scale as the stronger magnet. The attraction is not merely autonomy but the ability to deploy a billion-dollar check into a great company—something an independent manager can now contemplate while retaining the economics and avoiding partnership constraints.
3. Elad Gil exposes the difference between LP doctrine and LP behavior
Jason’s framing for LPs: they spent 20 years preaching focus, small funds, cohesive partnerships and board-level value-add, then financed a product that ignores each prescription. The message from successful managers is effectively, “Thank you for your advice. I’m ignoring it…now, thank you for your money.”
The charitable interpretation is that “idiosyncratic success triumphs over bland, mediocre, standard advice.” Elad’s seed exposure to Airbnb and Stripe, access to the best deals and sustained picking record give LPs a rational reason to discard pattern-matched rules: “Winners win.”
Harry compared Elad with a master property developer buying the best house on the best block. Whether the category is fintech, data or AI, the objective is to own the leader—examples included Stripe, Databricks, OpenAI, Harvey and Helsing—and own it meaningfully.
Jason kept the cycle-level warning intact. Some heuristics will prove outdated, but others may encode lessons about focus and key-person risk that only become visible during a downturn; virtually nothing founded after about 2010 has yet generated evidence across a complete cycle.
4. Deployment velocity buys relevance before it proves investment skill
Harry worried that “speed of deployment equals relevance,” potentially disadvantaging managers who maintain three-year fund cycles and temporal diversification. A fast investor stays in founders’ consideration sets and keeps receiving the information, access and markups that accompany a hot hand.
Jason reduced the mechanism to two outcomes: good rapid deployment creates “relevance in the short term and returns in the long term”; bad deployment creates “relevance in the short term and failure in the long term.” LPs alone must distinguish durable judgment from a feedback loop built on recent marks.
Tiger and SoftBank were the counterexamples. Their 2021 checkbooks produced universal access, but every capable early-stage investor knew which late-stage capital was “dumb money”; when selections failed, relevance evaporated. Elad represents the opposite case—a 10-to-15-year record spanning early and late investments that earns a longer runway.
Harry pushed back that Tiger did not simply pick badly; it “picked everything,” including positions such as OpenAI and Scale that may still rescue substantial value. Harry’s refinement was that Tiger played the available field: when the average public SaaS company grew about 70%, lowering the bar across triple-triple-double-double businesses looked less reckless than it does retrospectively.
5. One-person leadership scales better in late-stage investing than at Series A
Harry identified Thrive, Greenoaks, Founders Fund and Elad Gil among recent breakout managers and noted that all are dominantly one-person-led, even with strong teams beneath them. That is a meaningful break from the old venture partnership ideal.
Jason argued that early-stage ownership makes equality functional, not cultural. A partner taking a 10%–20% ownership position and speaking for the firm needs peers of comparable authority managing adjacent positions; Benchmark internalized this because one dominant partner surrounded by “betas” would be a poor steward of concentrated early-stage money.
Late-stage investing permits hierarchy because the work is primarily decision-making rather than ongoing company service. Deploying $1 billion in $20 million checks requires 50 decisions; using $100 million checks requires ten, making a single exceptional decision-maker far more scalable.
The cycle risk is correspondingly concentrated. Late-stage investors principally take valuation risk, and when that fails it can fail “in a 100% correlated fashion.” Rory would preserve his early-stage partnership model, though he conceded that had he known in 2009 that 15 years of rising equities lay ahead, a Tiger-like strategy—without the 2021 excess—might have been more lucrative.
6. The best late-stage franchises begin with early-stage access
Rory distinguished the named firms from financiers who arrive from public markets and treat venture as a spreadsheet exercise. The Facebook investment for $500,000 is the extreme specimen: an implicit $200 million–$300 million early-stage franchise can become a $3 billion platform whose pooled returns are driven by large late-stage checks.
The monetization loop is powerful: originate or access a company with a $10 million investment, then follow with $200 million at Series E. That lets an investor collect returns across the curve while using early-stage credibility as the entry ticket to late-stage scale.
Rory’s examples were Josh and Neil from Thrive and Greenoaks making more than $500 million from buying Carvana in the public-market collapse, with Neil also finding Windsurf at seed. Harry’s reaction to the falling-knife bet captured the point: “I give up. Neil—too good.”
7. Anthropic’s reacceleration is what the new valuation is buying
The round reportedly began around a $100 billion valuation, drew enough demand to rise toward $150 billion–$180 billion and, Rory heard, likely landed nearer the upper end. Against OpenAI around $300 billion, the move represented an immediate mark-up for investors who had preferred Anthropic at the prior price.
Rory focused on derivatives, not the absolute revenue figure: roughly $1 billion late last year and $4 billion this year would mean growth accelerating at scale, instead of decaying from 300% to 200% to 100%. “These guys appear to have reaccelerated at scale,” an exceptional signal if the numbers are correct.
Jason called the market structure “clearly a two-horse race, brutally,” with others either failing to ship, failing to monetize or remaining too small to matter. Anthropic is therefore being priced as the reaccelerating winner rather than merely another foundation-model laboratory.
8. Developer AI spend could move from $200 to $10,000 per month
Jason’s “Captain Obvious” realization came from heading toward an $8,000 monthly vibe-coding bill himself. With agents thinking for 15 minutes on debugging and complex tasks, he wanted four—and potentially 15—running concurrently; a nominally cheaper token becomes irrelevant if consumption expands faster.
Jason cited Farhan C—whom he thought was a CTO at Shopify—as an enterprise specimen: developers were allowed to use AI without caps, and the heaviest users consumed about $10,000 monthly by operating parallel processes. Jason’s endpoint is that every strong developer at a leading technology company receives $8,000–$10,000 monthly because “your agents don’t sleep.”
Harry challenged the call with Moore’s law: falling inference cost should reduce spending per developer. Jason’s rebuttal was that products improve exponentially, context and task scope expand, and agents can debug, refactor and build algorithms around the clock; demand may eventually plateau, but he sees no limit today.
The budget comparison drives the thesis. Spending $10,000 monthly is modest beside a $500,000 fully burdened developer who might leave in seven months, especially when top talent remains scarce even at Cursor and Lovable. Jason therefore sees developer-model TAM as perhaps 50 times current levels, not merely another $4-per-seat Jira product.
9. Usage caps are a bullish demand signal, even if margins are temporarily ugly
Harry interpreted Anthropic and Cursor limiting $200 plans as proof of “infinite demand” at a price that does not yet cover unconstrained usage. A provider might collect $200 while incurring $250 of tokens, cap use near $170, and tolerate poor point-in-time economics because compute costs keep falling.
The expected repair comes from both sides: next year the same capped $170 workload may cost $120, while the provider can raise the $200 price toward $400. The immediate result is pushing and shoving around gross margins; the durable result is a customer becoming more dependent as the underlying cost curve improves.
Harry put a possible $2 trillion Anthropic outcome to Jason. He declined to endorse that: trillion-dollar companies generally touch billions of consumers, while NVIDIA is the exceptional developer-adjacent case because it holds near-monopoly exposure to the most valuable commodity. He still expects Anthropic to become “extraordinarily valuable.”
10. Microsoft’s developer-tool lead became its most avoidable AI loss
Microsoft’s early Lovable/Replit competitor exposed the haste: in beta, users apparently shared one database and received warnings about what they stored. Jason read that less as serious execution than panic after Replit and Lovable approached $200 million in six months.
Harry’s sharper comparison was Cursor. Microsoft already owned GitHub, shipped Copilot roughly five years earlier and still reportedly sat near $500 million while Cursor reached about $900 million. The failure was especially stark because Microsoft had started earlier.
The failure complicates the “Microsoft is amazing in AI” narrative. It lacks its own frontier model, has a tortured relationship with OpenAI and is chasing another shiny object while its core developer franchise loses ground. Sadia wanted to “make Google dance”; Jason’s verdict was that Google now appears to be dancing very well.
11. Google leads the hyperscalers, though insatiable demand weakens the short
Forced to buy one hyperscaler and short another, Jason provisionally chose Google over Amazon. Five years ago AWS owned the category, Azure used enterprise relationships to reach second place and Google Cloud looked expendable; now Google is the smallest but “by far the fastest-growing,” while AWS appears the underperformer.
Jason accepted the execution ranking but would not place the short. Google owns TPUs, Amazon is already operating under “code red,” and every provider can sell more infrastructure than it has capacity to deliver. Harry’s observation that Oracle is suddenly competitive reinforces that the constraint is supply, not a lack of customers.
Harry’s concern is the gap between perhaps $600 billion of capex and a roughly $30 billion growing revenue line. Enterprises can only digest so much software annually, while hyperscalers must account for large infrastructure depreciation through earnings; Microsoft’s layoffs may partly defend efficiency as that charge reaches the income statement.
Timing makes the concern nearly untradeable. David Khan from Sequoas’s warning may ultimately prove right, but a trader expressing it early would have shorted Nvidia around $100 and watched it reach $170 while hyperscaler budgets rose. “The market can stay irrational longer than you can stay solvent.”
12. The incumbents are oligopolists—and still scrambling from behind
Harry rejected the claim that the giants are too powerful, preferring market correction to “idiot regulation.” The relevant AI markets contain three or four serious players—not a settled monopoly—including OpenAI, Anthropic, Google, Oracle, Microsoft, Amazon and others across the different segments.
His case for oligopoly is innovation: such markets compete on features more than price, keep products similarly priced and recycle rents into R&D. “If oligopolies aren’t okay, I quit this venture”; three or four technologically aggressive firms may be precisely the structure investors need.
Harry distinguished their old monopolies—Apple in iPhone, Google in search, Meta in social and Microsoft in corporate software—from the new contest. OpenAI and Anthropic are the entrants making the running despite incumbent balance sheets, which weakens the case for preventive intervention.
Harry’s memorable scorecard: Google “has executed the best of the four” because it has a working model; Apple has no working product; Microsoft bought someone else’s product but does not quite own it; and Meta is buying talent from “terrible psychological need.” They are “rich people on the back foot behind the new trend.”
13. Apple can survive an AI miss, but management may not deserve another cycle
Harry’s proposed Apple intervention was organizational, not acquisitive: the board should ask whether the management team is too old to seize AI. Apple is his largest position and has rewarded him for 15 years, yet “you haven’t grown in five years”; dividends and financial engineering do not substitute for winning the next platform.
Jason’s pushback was that AI may strengthen Apple’s tollbooth. He cited the App Store as roughly 25% of revenue and 40% of profit, with iPhone plus App Store around 75% of revenue; if more AI applications flow through iOS and Apple keeps about 26%, failing to build the model could still be lucrative.
Harry accepted that hardware has long inertia and makes the miss less existential than losing search would be for Google. The unresolved danger is an OpenAI-like personal companion coming between Apple and its billion users; monetizing product placement is defensible, but Jason’s warning was, “If you’re not moving forward, you’re dying.”
14. Meta’s control structure lets Zuckerberg spend through another mulligan
Zuckerberg’s talent siege is economically coherent if Meta spends $40 billion–$50 billion on capex: paying $1 billion–$5 billion for the few people capable of allocating it is not obviously excessive. Harry expects the marginal return eventually to turn negative, but admitted he had “no clue” whether Meta is 5% or 80% through the campaign.
Jason emphasized reversibility. Meta can stop, write off the contracts, harvest advertising and become a cash cow; a company producing roughly $40 billion–$50 billion of free cash flow can absorb “a mulligan at least every eight quarters,” as it effectively did with virtual reality.
The stronger advantage is that Meta and Oracle retain control. Ordinary profitable companies are trapped by EPS expectations and can scarcely reinvest without attracting punishment; these companies can telegraph a generational bet, spend otherwise trapped cash and remain protected from the vultures even if the stock falls.
15. Figma’s IPO pays investors after the zeitgeist has moved on
Jason expected the IPO script to work as designed: an initial filing range around $24–$28 moved to roughly $32–$35 after demand proved multiple times oversubscribed, making a high-end price and first-day pop difficult to avoid. Four venture firms each realizing about $1 billion would restore attention quickly.
Harry contrasted that with September 2022, when Adobe’s bid at 2021 pricing stunned 10,000 TechCrunch Disrupt attendees after the bubble had burst. Figma remains an “S-tier” software company, but Claude credits, Lovable and vibe coding now dominate imagination; traditional B2B software may simply be boring again.
Harry’s structural observation was harsher: “Our holding period is now longer than the tech cycle.” Figma took roughly 12–13 years to reach market and must already explain AI adaptation in its S-1; several pre-AI companies can still create huge public returns even after venture formation in their category stops.
Figma Make could turn the threat into a design-to-code war. Jason can “smell” a Lovable or Replit app within six seconds because pixel fidelity remains weak, yet he also hates static Figma links that do not work; the winning product makes the pixel-perfect prototype functional immediately. Today’s integrations are partnerships, but the panel doubts they remain friendly in 12 months.
16. The forecasts favor revenue momentum—and sovereign-funded valuation momentum
Jason initially said under, but accepted that the cited velocity and at least doubled developer spend could take Cursor from roughly $1 billion today toward $4 billion by the end of next year. Additional model usage would send more revenue toward Anthropic.
Lovable above $400 million ARR drew near-unanimous confidence after the cited 1-to-100 move in six months. Jason called missing the target a “colossal” failure, though he warned that valuation multiples could compress enough to produce a down round anyway; Harry identified churn and subscription duration as the real bear case.
OpenAI above an $800 billion valuation split them. Harry leaned under because successive funding step-ups are shrinking; Jason leaned over because GPT-5, Codex and a first Jony Ive product could renew excitement. Jason’s argument moved Harry somewhat: if OpenAI needs $800 billion, stakeholders, sovereign funds, warrants and structured terms can “solve for it,” even if public markets would not.
The outer bound is macroeconomic. AI data-center capex was cited around 1.2% of GDP, above the dot-com bandwidth peak near 1.1% but below railroads around 6%; unlike rails lasting 150 years, the transcript described the relevant asset as one that “appreciates over 3 years.” That comparison supports both camps: spending can still multiply, yet the bandwidth, railroad and Apollo comparisons all ended in sharp pullbacks.