🇧🇷 15y Public Equities (ex-Head of Research @ top Brazil fund). Experience w/ Alternatives & Illiquids. Former Public Co. Board Member. Tech enthusiast.
O risco estrutural de robotáxis (AVs = Autonomous Vehicles) para $UBER reside em market share incremental e compressão do take rate.
Minha conta de padeiro:
• Impacto esperado: −US$ 1,24 bi/ano de EBIT em regime (~−US$ 7,6/ação)
• 3/4 disso vem de corrida que sai da Uber e vai pro app da Waymo/Tesla
• Take rate menor é menos relevante
• O impacto se materializa mais fortemente em 2029/2030, à medida que a penetração aumenta
Passo 1: aonde está concentrado o impacto potencial?
12 cidades (SF, LA, Phoenix, Austin, Atlanta, Miami, Dallas, Houston, Orlando, San Antonio, Nashville e DC):
• ~26% do GB de Mobility (US$ 30,7 bi)
• ~31% do lucro do segmento (US$ 2,9 bi de US$ 9,5 bi)
A abertura por cidade é estimativa minha. A Uber só reporta a margem global (8,0%).
Passo 2: a curva de adoção
Hoje: ~6,2 mil AVs nos EUA, ~110 mil corridas/dia, ~1% do mercado.
Waymo: ~4 mil carros → ~8 mil no fim do ano → ~20 mil no fim de 2027 (fonte: Semafor).
Caso-base de 2030 (Reiner): 225 mil AVs, ~25% das corridas.
Anedota: em menos de 6 meses a $TSLA já tem o mesmo número de AVs rodando em Austin (1.3k)
Passo 3: os três vazamentos.
Quando o AV chega numa dessas cidades, a Uber perde de três jeitos:
• Direto: a corrida vai pro app da Waymo/Tesla. Some a margem inteira (~9,5% do GB), mais ~1,5% de custo fixo que fica.
• Hospedado: o AV roda na Uber, mas paga take menor.
• Promo: a Uber gasta pra segurar cliente e motorista.
Passo 4: quanto o dono do AV aceita pagar?
Ele só paga pela utilização extra que a Uber entrega. Em mercado maduro: 22,5 corridas/carro/dia na Uber vs 18,5 no app próprio.
Teto teórico: 1 − 18,5/22,5 = 17,8%.
O take rate de uma corrida tradicional com motorista paga 28,5%.
Passo 5: o detalhe que o consenso ainda ignora
Cada corrida custa ~US$ 4,96 ao dono (energia, seguro, manutenção, limpeza, teleoperação) numa tarifa de US$ 14. Ele divide a margem, não a receita:
17,8% × (1 − 4,96/14) = 11,5%
Mesmo com os +30% de utilização que a Uber alega: 14,9%.
Passo 6: a visão por corrida
• Humano: US$ 17,47 × 28,5% = US$ 4,98 de receita → US$ 1,66 de lucro
• AV a 18% e US$ 14: US$ 2,52 − US$ 0,85 de custo = US$ 1,67 (paridade)
• AV no teto de 14,9%: US$ 1,24 (−25%)
• AV a 14%: US$ 1,11 (−33%)
Passo 7: o P&L consolidado
Premissa: 20–35% do GB desses metrôs vai direto pros apps da Waymo/Tesla. Para cada 1 ponto que sai, ~0,65 ponto vira AV dentro da Uber, com take no teto. O resto segue no modelo legado, com promoções para reter os clientes.
Contra ~US$ 11,0 bi de EBIT ajustado FY27E:
• Bull: −US$ 0,72 a −1,25 bi (−6,5% a −11,3%)
• Base: −US$ 0,85 a −1,37 bi (−7,7% a −12,4%)
• Disrupção: −US$ 1,53 a −2,02 bi (−13,9% a −18,3%)
Passo 8: o impacto esperado
Pesos de 30% Bull, 45% Base e 25% "Disrupção", com 27,5% indo direto:
• Direto: −US$ 0,93 bi (75%)
• Promo: −US$ 0,21 bi (17%)
• Take menor no AV hospedado: −US$ 0,10 bi (8%)
Total: −US$ 1,24 bi, ou −11,2% do EBIT ajustado (−13,8% do GAAP). Com 21% de IR, −US$ 0,48 de EPS.
A 16x, −US$ 7,6/ação: US$ 70,55 → ~US$ 62,9.
Passo 9: a velocidade que o impacto chega
Os −US$ 1,24 bi são o cenário materializado, não 2027e. No caso-base, o lucro de Mobility contra um mundo sem AV:
2026: −0,5%
2027: −1,2% (−US$ 0,12 bi)
2028: −2,6%
2029: −5,8%
2030: −12,6% (−US$ 1,75 bi)
O take rate médio vai de 25,3% para 24,2%. No cenário de "Disrupção", o dano chega a −20% em 2030.
Passo 10: os catalisadores para monitorar
• Fim de 2027: Waymo com ~20 mil carros; Tesla com ~5 mil (estimativa do Reiner); Uber com 2.000 AVs
• Jan/2028: acaba a exclusividade da Waymo com a Uber em Austin e Atlanta (segundo a imprensa; a Uber não confirmou)
• 2028+: Uber mira 10 mil+ AVs, com capital de terceiros
Passo 11: o $UBER investe em frota própria
Depreciação não é o problema: US$ 163–200 mi/ano para 10 mil carros, ou 1,5–1,8% do EBIT.
A conta certa é cash on cash: caixa por carro ÷ capital bruto (US$ 75 mil do carro + US$ 15 mil de depósito = US$ 90 mil).
A US$ 14 a corrida:
• US$ 14 − US$ 4,96 de custo variável − US$ 0,85 de plataforma = US$ 8,19 de caixa por corrida
• × 22,5 corridas/dia × 365 × 92,5% de disponibilidade (parado 7,5% do tempo) = ~US$ 62 mil/ano por carro
• ÷ US$ 90 mil de capital (US$ 75 mil do carro + US$ 15 mil de depósito) = 69% de cash on cash antes de IR. Payback de ~1,4 ano (~1,7 depois de IR).
Só que o carro dura ~4 anos, e a curva real de depreciação ainda é uma incógnita. Com US$ 10 mil de valor residual e o depósito perdido, é preciso ~31% de cash on cash só para render 10% de TIR. A US$ 14, a TIR é de ~48%.
E retorno de elevado atrai concorrência, que derruba a tarifa:
A tarifa de equilíbrio (IRR de 10%) fica em ~US$ 9,6–11,5, versus o $14 atual. O risco é tarifa e capital (US$ 900 mi por 10 mil carros, 2,5x o capex LTM), não EBIT.
O melhor contra-argumento até agora:
• O CFO diz que Waymo e Tesla fazem "low teens" corridas/carro/dia nos mercados novos, contra "high 20s" na Uber. Se isso valer em escala, o teto sobe para 21–32%.
• O mercado cresce ~13% a.a., até ~17,7 mi corridas/dia em 2030 (Reiner). A Uber cresce ~10% a.a. mesmo perdendo share.
• Se o app próprio pagar CAC de ~US$ 1,25/corrida (hipótese minha), o teto de 11,5% sobe para ~18,8%.
Ou seja, o risco seria de lucro incremental e múltiplo que o investidor topa pagar, mas não de existência.
A ação ~US$ 62,9 fica perto do meu preço de entrada (≤ US$ 60, 13,1–13,6x FY27E, TIR ≥15% em 3 anos). São contas diferentes: os US$ 60 já descontam um cenário de AV, e os US$ 62,9 são só o ajuste estático deste modelo.
Kill switches (cada um basta sozinho):
• Share direto de Waymo/Tesla >15% em 2+ dos 5 maiores metrôs antes de 2029 (o caso-base só chega lá entre 2029 e 2030)
• Capex + leasing de AV >US$ 1,0 bi/ano, ou dívida líquida/EBITDA >2,5x
• Take híbrido reportado <11,5%: a Uber não teria poder de barganha nenhum
O que me prova errado do outro lado:
Take rate ≥18% reportado em frotas Gen-6 de terceiros, ou AVs rodando 30+ corridas/dia na Uber em mercado maduro.
Nesse caso tiro o cap e corto a probabilidade de "Disrupção" para 10%. O dano esperado cai para ~−US$ 0,94 bi (~US$ 64,8/ação).
I was in Austin this week and only used Uber to go to the airport. Cybercabs are way superior. Prices were either equal or better than Uber, and waiting time was fine, some times even even quicker than Uber, when not, I’d rather plan ahead and call Cybercab and have the car waiting for me outside (7min - no charge).
A arquitetura corporativa que Elon Musk opera como um sistema fechado de criação e monetização de gargalos. Steve Davis, CEO da Boring Company, resumiu essa lógica com clareza.
Os robotáxis já operam comercialmente em cerca de 12 grandes cidades americanas. À medida que a frota escala e o custo por km desaba, o volume de deslocamentos dispara e multiplica o número de carros nas ruas. O sucesso da Tesla cria imediatamente o próximo gargalo físico: a saturação severa da malha viária.
Escalar o transporte e a infraestrutura além do limite bidimensional (2D) das cidades exige ocupar o eixo vertical também (3D):
- Subsolo (The Boring Company): A escavação tradicional de túneis urbanos nos EUA consome entre US$ 500 milhões e US$ 1 bilhão por milha (~1,6 km). Ao cortar esse capex em mais de 90%, para a faixa de US$ 10M a US$ 50M por milha, a Boring Company viabiliza artérias subterrâneas de alta vazão para escoar o fluxo autônomo.
- Órbita e Espaço (SpaceX): Onde a superfície terrestre atinge limites físicos e energéticos, a SpaceX estende a infraestrutura de conectividade, transporte de longa distância e capacidade computacional para fora do planeta.
Cada companhia do portfólio monetiza a externalidade gerada pela outra: a Tesla satura a superfície, a Boring Company escava a válvula de escape subterrânea e a SpaceX captura a camada orbital.
$TSLA
$SPCX
Aravind Srinivas (Perplexity founder) said he’s glad he doesn’t have the difficult job of trying to keep Google Search a growing business.
“The interface only has so much real estate.”
With AI Mode delivering direct answers, users have less reason to click through and browse further. The challenge grows even larger once AI agents become the dominant internet users.
I still believe Google will find success monetizing the AI Mode surface (perhaps temporary, like Meta’s success with Reels). Time will tell.
$GOOGL
Another evidence of plateauing CapEx.
In my view, this is bullish the CSPs (ROI will show) / app layer (I believe AI is a really transformational tech and it’s only a matter of time to new use cases to show), and bearish semis.
$Tencent / Chinese CSPs’ tone on the AI capex cycle is very clear -
Martin Lau laid out a clear case for why AI capex should moderate after 2027:
1/ Front-loaded investment: AI infrastructure is like building a factory. Once built, it can continuously train new models and generate tokens.
2/ Lower hardware costs: The same equipment should cost significantly less by 2028.
3/ Rapid efficiency gains: The same infrastructure can deliver exponentially more training compute and token output over time.
Takeaway: Martin is effectively calling for a peak in AI capex in 2026-27, with a meaningful decline in capex intensity thereafter, even as AI demand continues to grow.
The read-through for semis is interesting: If annual industry AI capex eventually normalizes below $1T, assigning a mid-range multiple to normalized earnings, that’s your industry total size.
Chip financing as an asset class: both equity and debt will finance $1.5T (60–80% loan-to-cost) over the next 1-2 years.
This is going to further pressure credit spreads.
As spreads creep higher, prepayments will play a bigger role in financing the buildout, yielding more concentration among the big-balance-sheet guys.
Some perspective from Chinese AI Labs regarding the Anthropic IPO.
https://t.co/ykWM7okxZy combined with Minimax are worth less than US$30B. Revenues amount to ~US$1.3B sales, while consensus estimate +200% sales growth for 27e.
EV/Sales (NTM)
- Minimax 4.7x
- https://t.co/NDiexYEIpB 10.5x
Both are down 60%-90% from peak.
Feedback from the ground: Chinese AI labs are increasingly desperate as access to frontier models for distillation gets cut off. Asian investors are shorting Chinese CSPs (facing capex creep) and AI labs while buying local semiconductor suppliers in the wake of the latest US export controls.
This should be a tailwind for Anthropic’s ARR over the coming quarters. Distillation was a real factor.
@TomorrowXSummit
@RagingVentures It'd make a lot of sense, especially now that we have one player less financing the DC buildout, namely $NVDA. Additionally, it seems Dario believes securing capacity equals competitive advantage in the long term.
@TomorrowXSummit Day 1 Scratchpad: Supply Velocity vs. Token Demand
Builders radiate conviction while investors sit defensively on the sidelines. Most have largely pulled chips off the table, mirroring Druckenmiller’s warning.
Underwriting the supply side handed the Street an effortless trade over the past two years: hyperscaler CapEx guidance translated mechanically into semiconductor top-line beats. Software engineering delivered the first mass-market token sink because developers sit directly inside the toolchain. That initial adoption wave is now past us.
The market requires a second mass-market use case to underwrite forward token demand. Consensus obsesses over datacenter power capacity in gigawatts, yet the critical variable is token supply, which compounds 3x every single year.
Cybersecurity possesses every catalyst to emerge as the next structural token sink. Frontier models have exposed systemic vulnerabilities across enterprise and sovereign infrastructure, elevating AI deployment safety to a national-security mandate; punctuated by last week's operational disruptions across US airports and major banks.
Cybersecurity equities already embed elevated growth expectations in their multiples. Absorbing inference compute into cost-to-serve will compress gross margins, yet accelerating top-line expansion should outpace the margin haircut and grow gross profit dollars. Most importantly, an enterprise security inference cycle de-risks the demand side of the AI trade over the next twelve months.
Beyond security, enterprise operators must ship native workflows that replace the raw prompt box. The builders' technical conviction is grounded in reality: foundation models can generate nearly any digital artifact and unlock massive creative output. Commercial diffusion accelerates once software wraps that raw intelligence into deterministic workflows.
Long-term economic utility will materialize, yet portfolio positioning right now demands clean asymmetry.
Hyperscale CSPs win across every trajectory because compute scale is a balance-sheet game, and enterprise AI workloads run directly on their cloud rails:
A. If neoclouds crack: CSPs deploy fortress balance sheets to acquire stranded capacity at distressed cents on the dollar.
B. If CapEx pauses: Free cash flow inflects sharply upward, handing skeptical allocators the immediate cash ROI they demand.
Trading at palatable multiples, CSPs capture structural AI upside while insulating the book against a second-derivative hardware digestion phase.
@Jason After that WSJ article that exposed how Colette was more conservative. After all, she is the CFO signing that balance sheet.
I expected the company would raise buybacks as well, especially now that the stock collapsed.
https://t.co/nO8HKsTWcY
Many reached out to discuss robinhood:0xd0601ce157db5bdc3162bbac2a2c8af5320d9eec. Here are my thoughts.
Portfolio BLUF: The structural compounder thesis has expired (company was over-earning and normalizing EPS gets ~30% cut in my napkin math), leaving only a tactical normalization trade on the table. robinhood:0xd0601ce157db5bdc3162bbac2a2c8af5320d9eec should recoup the "Jensen drawdown" as Q3 prints clear near-term air pockets and restore visibility. With Colette’s more conservative hand on the wheel, expect management to step up share buyback, putting a hard technical floor under the tape and giving institutional holders cover.
The AI trade as we knew it is over. And it ended with Jensen.
Jensen extracted peak scarcity rents at 70%+ gross margins and weaponized allocation control to stretch the runway. Those fat margins gave hyperscalers urgent economic incentive to accelerate custom ASICs so they could take back cost and architectural control.
To dilute hyperscaler concentration and keep volumes moving while custom silicon ramped, NVIDIA pushed supply into sovereign builds and neoclouds, widening the CUDA moat. At the same time, management promoted open-weight models and launched Nemotron to seed ecosystem demand and reduce reliance on any single frontier lab.
Every one of those moves made tactical sense on the board, yet each carried steep trade-offs. Stuffing lower-tier channels and underwriting ecosystem demand with NVIDIA’s own balance sheet protected near-term top-line growth and foundry priority at the direct expense of revenue quality.
What the Street priced as an unassailable moat (total supply-chain control) became an operational trap. Owning allocation forced NVIDIA to clear silicon to any marginal buyer capable of keeping the flywheel spinning: vendor-financing customers, cutting spot deals, and feeding balance-sheet-light neoclouds. Shipment velocity beat counterparty credit quality every time. Any successor inherits a decelerating capital cycle stripped of the founder-level force that ran that machine at max speed.
The cash-flow statement has been flashing red for quarters. Cash margins are compressing alongside worsening revenue quality. Cash conversion broke down hard in the latest print, as operating cash flow lagged net income as management stretched payment terms, ballooned accounts receivable and DSOs, and poured free cash flow into ecosystem equity stakes. Layer worsening cash conversion onto a multi-year datacenter duration profile, and multiple compression follows naturally.
Colette’s public and private pushback on aggressive channel expansion told you everything: management was already trading away earnings quality to prop up headline growth.
Now NVIDIA is boxed into a structural lose-lose. Keep pushing GPUs into neoclouds, and you hit a credit wall the moment private funding dries up or end-user utilization stalls. Rotate that supply back to hyperscalers, and you have to grant the exact pricing discounts and commercial concessions Jensen refused to give. Both paths compress gross margins and break the growth narrative holding up the multiple.
That secondary-channel scaffolding also creates direct upstream contagion. NVIDIA deployed take-or-pay capacity floors, lease and power guarantees, equity checks, and stretched terms so weak credits could absorb silicon their own balance sheets lacked the capacity to fund. A material slice of the order book and capacity forecast NVIDIA hands to TSMC rests directly on self-underwritten balance-sheet engineering.
Once more conservative leadership tightens credit, or neocloud utilization disappoints, the demand signal backing NVIDIA’s dominant CoWoS and leading-edge wafer allocations softens immediately. Any order pacing or wafer reallocation at $TSM will expose how much of the prior cycle was vendor-financed pull-forward.
Late-cycle balance-sheet degradation was the price of keeping the flywheel spinning once Tier-1 hyperscaler demand plateaued against supply. The regime has turned. The trade is over.
“Here’s my prediction: the cost of AI inference per task will reduce by a factor of 100x in the next 2 years. But usage will grow more than one thousand times.” - Lin Qiao, Fireworks