Inder's Desk: Research Notebook

Google Is Becoming a Compute-First Company

Inder Sabharwal — Sun, 09 Aug 2026 06:50:19 GMT

What the company does and how it makes money

Q2 2026: revenue $119.8B, up 24%; operating income $40.8B at a 34% margin. The segment chart above shows where both come from.
Search & other is the engine: $63.3B, up 16.8%, the largest single line by far.
The rest of ads: YouTube $11.1B, up 12.9%; Google Network (ads on third-party sites) $7.3B, down 0.7%, the only revenue line that shrank.
Subscriptions, platforms and devices: $12.9B, up 15.2%. One blended line: YouTube TV/Music/Premium, NFL Sunday Ticket, Google One, Play purchases, and hardware, including publicly known products such as Pixel, Fitbit, and Nest (the filing itself just says “devices”). Alphabet discloses no split, so nobody outside the company knows what YouTube subscriptions or Pixel actually earn.

Google Cloud: $24.8B, up 82%, the fastest-growing line. The segment also includes Workspace and, new this quarter, TPU hardware sales.
Other Bets: $382M of revenue on a $1.8B loss that widened 45%. Revenue comes from Waymo rides and internet service (likely Google Fiber, not named in the filing); a third-party estimate below suggests internet service, not Waymo, is actually the larger piece.
The profit mix the revenue lines hide: Services earns $39.5B of operating income, roughly 97% of Alphabet’s total and 4.5x Cloud.
Waymo, the flagship bet: ~400K paid weekly rides in early 2026, ~500K by mid-year, targeting 1M by year-end.
As of May 2026, Waymo’s coverage had passed 1,400 square miles across 11 cities, with ~20 planned including London and Tokyo (estimated revenue split vs. the rest of Other Bets in the chart below).

Waymo’s February raise looks bigger than it is: $16B at a $126B valuation, but $13B of it was Alphabet’s own money. Only ~$3B came from outside, led by new investors Sequoia, DST Global, and Dragoneer alongside existing backers including a16z and Mubadala, so the price is market-set only at the margin.

GCP vs. the Gemini frontier lab: two different businesses, one shared bill

Only one of these has a separately disclosed P&L. That asymmetry is the story.

Cloud’s margin is now 35.6%, up from 20.7% a year ago. AWS grew 28% and Azure 40% in their latest quarters; Google is running at roughly double both.
The backlog hit $514B, up $50B+ in a single quarter; management expects half to convert within 24 months. ~90% of the Fortune 100 are on Gemini Enterprise.
Still the junior engine: Cloud’s $8.8B of segment profit is under a quarter of Services’ $39.5B.
Gemini/DeepMind has no disclosed P&L at all. Some of its shared costs land in “Alphabet-level activities,” a bucket that also holds philanthropy, corporate overhead, and fines; the bucket’s loss widened from $3.4B to $5.8B. How much of that is the lab? Alphabet doesn’t say.
Cloud’s 35.6% margin also excludes whatever share of that shared AI research economically belongs to it; the filing doesn’t allocate.
SemiAnalysis (”Gemini is Cooked but GCP is Cooking,” Aug 7) argues Google is deliberately trading frontier competitiveness for near-term Cloud monetization. Their thesis, not Google’s words.
Their sharpest claim: more than 20% of TPU shipments from late 2026 through 2027 are going directly to Anthropic, Google’s most direct rival lab. The 20% is their supply-chain estimate; the deal itself is public.
They rank Gemini “8th or 9th” among frontier models and estimate ~$12B of Gemini ARR against ~$73B of third-party AI ARR for GCP by end-2027. Rankings and dollars are their model, not Alphabet disclosures.
Then the humans walked. Jeff Dean, Chief Scientist, employee #30, 27 years, announced his exit on Aug 5 to co-found Discovery Loop, an AI-for-science startup, alongside Sanjay Ghemawat, Oriol Vinyals, and Quoc Le. Friendly split; Google is a founding investor and Cloud partner.
Same announcement: Hassabis moved up to Alphabet Chief Scientist and DeepMind Chairman; Koray Kavukcuoglu took over day-to-day DeepMind operations.
The stock fell 4.03% that day (IBKR-verified), the only hard number the event has produced. Four senior departures announced together materially support the lab-is-slipping read; it’s also consistent with an amicable spin-out. Both readings stay open.
Management flagged two H2 Cloud margin headwinds (rented third-party capacity, Wiz integration) and twice declined to say whether TPU sales help or hurt segment margin.

Frontier Model Economics

GCP is growing fast, accelerating for five straight quarters to 82%. Gemini’s own token growth is doing the opposite: 60% in Q1’26, 38% in Q2’26, the only two quarters SemiAnalysis has estimated so far.
On Artificial Analysis’s capability index, Gemini 3.1 Pro Preview scores lowest of the eight rated models (48). Meta’s Muse Spark 1.2 (57) scores higher and costs less on both input and output; Alibaba’s Qwen3.8-Max (58) also scores higher, ties Gemini on input price, and costs half as much on output. Google isn’t just behind on capability; it isn’t winning on price either.

Search, Cloud, and where they land in two years

Cloud’s growth rate has climbed an average 10.8 points a quarter, 28% to 82%. Nothing else in this business moves like that.
Search’s growth roughly doubled over the same six quarters, 9.8% to 19.1%, then gave back 2.4 points in Q2 2026. Whether that matters is the next chart.
Gemini’s token growth is the only line falling, and the only one that is neither revenue nor an Alphabet figure.

Cloud’s band is one question: when the growth rate peaks. Three dates, because that is what the answer turns on. Peak at the end of 2026 puts Cloud at $55B a quarter by Q2 2028; end of 2027, $67B; end of 2028, meaning no peak inside the window at all, $73B.
Nothing says the peak has happened yet. Cloud’s rate is accelerating into the last print, by 14.2, 15.6 and 18.4 points across the last three quarters, and the current ascent has run 11 quarters from the 22.5% trough in Q3 2023. The prior decline took 9 quarters, and before that the rate sat above 42% for 6.
The end-2027 case shares that six-quarter count, though not its shape: historically the rate was merely above 42% for six quarters with the peak in the middle, while this scenario holds today’s rate for six more. An echo, not an equivalence.
Once it turns, all three glide at the pace this segment actually set last time, peak to trough: 53.9% in Q2 2021 down to 22.5% in Q3 2023, about 9.3% a quarter off the rate. Even the earliest peak leaves Cloud growing 40% in Q2 2028, roughly where Azure is now.
Search now carries a real bear case rather than a flat line. The upper edge holds today’s rate. The lower edge erodes it by twice the worst four-quarter deceleration Search has printed outside COVID, which turns growth negative from Q1 2028 and lands at -3.9%, with Search shrinking to $64B a quarter.
That stress case is a deliberate handicap, not a reading of the data, and the chart says so. Search’s broader trend improved through Q1 2026 and its two-year rate is roughly flat; only the latest one-year print stepped down. The case exists because the risk is real, not because the filings show it.
Put the two together and the crossover becomes conditional rather than distant. Cloud passes Search by Q2 2028 in the two later peak cases, but only against a stressed Search. Against a Search that simply holds its rate, Cloud reaches 87% of it and no further.
Year one is the firm part: Cloud adds $56B to $61B against Search’s $21B to $41B. Year two is softer and the chart marks where it starts, because from there the year-ago base is itself a projection and the arithmetic compounds on its own output.
TPU rides GCP’s rate throughout, reaching roughly $2.5B to $3.2B a quarter by Q2 2028 depending on the peak case. Management put the vast majority of it in 2027, faster than the segment around it, so it is likely understated. Inventory did jump from $2.4B to $10.0B in six months and management linked the build to TPU, but Alphabet discloses neither the TPU share of it nor the hardware margin, so no revenue capacity can honestly be derived from it.

Alphabet total revenue is $119.8B a quarter today. On the central case it reaches $193B by Q2 2028, with the outer cases at $158B and $198B, spanning the same uncertainty already shown for Search and Cloud, compounded.
The mix is the part worth watching. Cloud goes from 21% of the company to 35% on the central case, and it does that without Search shrinking in any case except the stress one.
Everything else, which is YouTube ads, Google Network, subscriptions, platforms, devices and Other Bets, is derived by subtracting Search and Cloud from filed total revenue. It is exact rather than estimated, but it is a composite of very different businesses held at one rate here.

None of the projections above capture what Alphabet owns. The balance sheet carries $225.6B of equity holdings that produce no revenue line at all.
SpaceX is named in the filing itself, which is unusual and makes it the hardest number here: $80.0B of shares under short-term sale restrictions, plus $14.1B restricted through the third quarter of 2027. That second date is the only disclosed unlock in the set.
The other $131.5B is non-marketable securities, filed but with no component named. It nearly doubled from $68.7B at year-end, which is where the quarter’s equity gain came from.
Anthropic appears nowhere in the 10-Q. Not once. The widely repeated claim that part of that gain is an Anthropic stake rests on a single hedged report, and if the stake exists it sits inside that unnamed block.
Waymo appears nowhere either, for the opposite reason: it is consolidated inside Other Bets, so it carries no asset value at all. Its last round valued it at $126B in February 2026. That number is in neither the revenue lines, where Waymo shows up as $382M a quarter against a $1.8B loss, nor the assets.
A holding is not cash, and this is where the optionality framing earns its caveat. The SpaceX tranches are restricted, private stakes are illiquid, and selling any of it is a decision Alphabet has not announced. What the filing establishes is that the option exists and roughly what it is worth, not that it will be exercised.

The bull case

Cloud is compounding: five straight quarters of accelerating growth, margin up 15 points YoY. Its latest quarter ran at roughly double AWS’s and Azure’s most recent results.
TPU hardware revenue has barely started; the “vast majority” lands in 2027. SemiAnalysis models 2027 Cloud growth in the mid-100s percent against a 64% Street consensus. And selling TPUs to Anthropic pays Alphabet whichever lab wins.
The funding engine is large: $39.5B a quarter of Services profit, ~$70B freshly raised, and capex guidance lifted to $195B-$205B for AI infrastructure.

The bear case

Cash quality: free cash flow went negative (-$5.9B) for the first time in this series, capex doubled year over year, and long-term debt roughly doubled to $98.2B.
EPS quality: $9.11 was 69% one-time gain. The $99.0B equity-securities markup, mostly Alphabet’s SpaceX stake after its June IPO, added $6.26; ex-gain, ~$2.85 against a ~$2.89 consensus.
One report, unconfirmed, attributes part of that gain to a stake in Anthropic: the same rival lab Google sells TPUs to.
The lab: a widening shared-cost bucket with no disclosed offset, a model one analyst shop ranks eighth or ninth, and four senior departures announced in a day.
Google Network declined 0.7% YoY in Q2, and Cloud margin carries two management-flagged second-half headwinds.

What settles it

Alphabet’s next quarterly report: does the Alphabet-level loss keep widening, does Cloud margin hold against the flagged headwinds, does the backlog build again.
Gemini 3.5 Pro’s launch reception versus SemiAnalysis’s pre-committed “roughly Opus 4.5 level” call. Independent benchmarks will make that checkable.
Waymo’s 1M weekly rides by end of 2026: dated, binary, hit or miss.

Appendix: sources

Alphabet Q2 2026 earnings release, 8-K exhibit 99.1, filed 2026-07-22, SEC EDGAR. Primary source for all revenue lines, segment operating income, segment definitions, and the $99.0B equity-securities gain.
Alphabet Form 10-K, FY2024, SEC EDGAR. Source for the Search & Other and Google Network revenue-line definitions (stable boilerplate language across recent fiscal years).
Alphabet Q2 2026 earnings call, 2026-07-22, referenced via SignalDeck’s prior verified GOOGL Cloud deep-dive. Call-only disclosures: the $514B backlog and its conversion expectation, the two H2 margin headwinds, the capex guidance raise.
SemiAnalysis, “Gemini is Cooked but GCP is Cooking,” 2026-08-07. Competitive rankings, Gemini ARR, TPU/Anthropic shipment share. All figures are SemiAnalysis’s own model, not Alphabet disclosures.
Artificial Analysis, model intelligence index and pricing, with individual model pages fetched directly for each score in the frontier-economics chart. Data note: Artificial Analysis’s live page for Meta’s Muse Spark 1.2 shows an Intelligence Index of 57, while a dated Artificial Analysis article shows 54 for the same model, likely an index-recalibration artifact, not resolved further; the chart uses the live value.
First-party pricing pages: Anthropic, OpenAI, Alibaba Cloud Model Studio, Google Gemini API, Moonshot AI. All fetched 2026-08-07.
Meta’s Muse Spark 1.2 API pricing: The Decoder (2026-07-09) announced Muse Spark 1.1 at $1.25/$4.25 per million tokens; OpenRouter’s Muse Spark 1.2 pricing page (checked 2026-08-07) independently lists the same $1.25/$4.25 for the 1.2 model specifically. OpenRouter is a third-party API marketplace, not Meta’s own pricing page; no first-party Meta pricing page was found for this model.
Jeff Dean departure coverage, 2026-08-05: SiliconANGLE, CNBC, Axios, GeekWire. Convergent reporting; no primary Google blog post was retrievable.
Waymo funding, 2026-02-02: Bloomberg, CNBC. The $16B round, $126B valuation, Alphabet’s $13B share.
Waymo ridership and expansion: Forbes (2025-12-10) on the 1M-weekly-rides target; Electrek (2026-05-13) on the >1,400 sq mile, 11-city coverage figure.
Sacra, Waymo revenue estimate, as of Feb 2026. Basis of the Waymo-vs-rest Other Bets split chart, itself an estimate, not an Alphabet disclosure.
AP/Fortune and Motley Fool, both 2026-07-23. Attribution of the equity gain to SpaceX; the Anthropic attribution is Motley Fool’s alone, hedged as “reported to be,” unconfirmed by the other outlet.
Interactive Brokers market data, verified 2026-08-06. GOOGL closes: $377.65 (8/4), $362.43 (8/5), the -4.03% announcement-day move.

Disclaimer:

This article is for informational and educational purposes only and does not constitute financial, investment, tax, or legal advice. Any opinions, scenarios, price targets, or market observations reflect my personal views and may change without notice. Investing and trading involve substantial risk, including the possible loss of principal. You are solely responsible for your own investment decisions, position sizing, risk management, and trades. Conduct your own research and consult a qualified professional where appropriate.

$SPCX: A Broadband and AI Compute Company With a Rocket Program Attached

Inder Sabharwal — Wed, 05 Aug 2026 17:15:53 GMT

Data as of the Q2 2026 10-Q and same-day earnings call, August 4, 2026. Call statements are management commentary, unaudited and not part of the filing. Not investment advice.

SpaceX just filed its first quarterly report as a public company, and it confirms what the merger chain (Twitter into xAI in March 2025, xAI into SpaceX in February 2026) already hinted at: the company everyone still thinks of as a rocket maker is now mostly a satellite broadband business bolted to an AI infrastructure buildout, with launch the smallest segment by revenue. The earnings call added the part the filing couldn’t: management telling us, in its own words, what the AI business is actually for.

What SpaceX Does and How It Makes Money

Three segments, one company. Space sells rocket launches and government development contracts. Connectivity sells Starlink satellite broadband to consumers, enterprises, and governments. AI is Grok, the X platform, and, increasingly, cloud compute rented out to customers.

The segment picture:

The profitability picture is the opposite of the public image: Connectivity is the only segment making money on GAAP operating income. AI lost money operationally but ran positive on Adjusted EBITDA. Space lost on both. Add it all up and the company still ran a $143M operating loss this quarter.

What’s Actually Driving the Revenue

Starlink is the engine. At 55% of revenue, Connectivity is the only segment profitable on GAAP operating income and the biggest Adjusted EBITDA contributor (+$2,597M of the company’s $3,538M). Enterprise & Government (+108.3% to $1,806M) is growing faster than Consumer (+44.4% to $2,485M).
The growth is volume, not price. Subscribers doubled to 12.0M, but ARPU fell 22.4% to $66/month from $85 as the mix shifted toward cheaper international and consumer plans – the funnel is widening even as each user pays less.
AI is the fastest grower but the most capex-hungry and still GAAP-unprofitable – more on that in a moment.
Space grew in Q2 but shrank over the half. Q2 revenue rose 29% (Shotwell cited $6B+ of new U.S. contracts won in the quarter [CALL]), but first-half revenue fell 1.9% (17 customer launches vs 21 a year ago), and the segment lost money on both operating income and Adjusted EBITDA.

The AI Business, In Full

What it is: Grok, the X platform (subscriptions and ads), and AI compute infrastructure, all inherited through the xAI merger. Nameplate compute reached 1.4 GW at quarter-end, 3.5x a year ago.

What’s growing: compute for rent, not Grok - Musk said on the call he expects “maybe 10%” of SpaceX’s compute to go to Grok training, with the rest split between running inference for SpaceX’s own products and renting capacity out [CALL] – he didn’t say how that remaining ~90% splits. - The Q2 numbers back him up: AI Solutions & Infrastructure revenue went from $311M to $2,194M, with ~$1,600M of the increase coming from cloud services the filing says SpaceX “began to offer” this year, versus $258M from Grok/X subscriptions. Advertising actually fell 13.9% to $367M during a platform transition. The CFO tied the $1.6B ramp to new cloud agreements at the Colossus and Colossus 2 sites [CALL]. - Training Grok is a minority use of the buildout. The revenue growth is selling capacity.

Where the AI revenue came from, this quarter vs. a year ago:

The customer concentration is extreme, straight from the filing: an unnamed “Customer B” is 19.5% of total revenue, AI-only; add “Customer A” (18.3%, spanning all three segments) and two customers are ~38% of the company. Management named Google and Anthropic on the call as AI customers [CALL], but the agreements they described ramp after June 30 – they tell you about the forward mix, not who was actually paying in Q2.
The contracts are also short-notice: a new risk factor [10-Q] discloses that cloud agreements are generally cancelable on 90 days’ notice after ramp. That’s a class-wide term, not specific to any customer, but this revenue isn’t locked in the way Space’s backlog is.
The loss-vs-EBITDA gap is mostly depreciation timing: the GAAP loss was $1,257M (better than $1,524M a year ago, though the first-half loss still widened 51.5%), while Adjusted EBITDA swung to +$1,146M from a $276M loss. AI took 86% of quarterly capex, and $12,554M of construction-in-progress isn’t depreciating yet. That’s a real signal, but not automatic convergence – new capacity brings new depreciation with it.
The forward bet is contractual: $27,955M in non-cancelable commitments, $22,244M of it due in 2027. And SpaceX is now, per Musk, exclusive to Nvidia [CALL] – one vendor on supply, concentrated customers on revenue.
The financing has a governance wrinkle: $13,329M of the buildout runs through “failed sale-leaseback” deals with Valor Equity Partners, whose founder sits on SpaceX’s board.
Beyond Grok: a $60B all-stock Cursor acquisition closes in Q3. Competitively, per dated reporting [3P]: Grok is #3 in US chatbot share (~17.8% as of January 2026, behind ChatGPT’s 52.9% and Gemini’s 29.4%), plateauing since March, smaller than Anthropic or OpenAI standalone.

What the AI buildout has spent, and what it’s still on the hook for:

The Bull and Bear Case

What Settles It

New named-scale AI customers beyond the confirmed Google/Anthropic ramps – broadening supports the compute thesis; another quarter of two-customer dependence weakens it.
AI’s GAAP loss trajectory as in-service assets catch up – movement toward positive EBITDA supports the depreciation story; a widening gap weakens it.
Capex against guidance: the CFO guided the next two quarters similar to Q2’s ~$18.4B, with a stated end-2026 target above 2 GW [CALL]. Landing near guidance with revenue attached supports the buildout math; slippage weakens it.
Near-term catalysts: the Cursor close (Q3 2026) and progress toward the final Spectrum Acquisition Closing (pending, 2027-2028).

Bottom Line

is now a profitable broadband segment, an AI land grab, and a rocket program in one ticker. Launch is 12% of revenue, and Connectivity is the only segment with positive GAAP operating income.
Starlink answers “what drives revenue”: 55% of sales, the largest share of segment EBITDA, growth from doubling subscribers while ARPU falls 22%.
The AI business is, by management’s own account, mostly not about training Grok: ~10% of compute for Grok training per Musk, the rest split between internal use and rental. The growth is cloud capacity, the bulk of the segment’s disclosed increase, sold under contracts generally cancelable on 90 days’ notice, with two customers at ~38% of revenue (the AI-only one plausibly Anthropic, the other unidentified), $22.2B committed for 2027, a single-vendor Nvidia supply chain, related-party financing, and a flagship datacenter under litigation – all while Grok sat third in reported US chatbot share as of January 2026.
The market’s excitement and the AI business overlap but aren’t identical, and its current preference is unsettled: $SPCX trades at a premium to a same-growth, GAAP-profitable comp unless you take management’s $100B ARR target at face value, and the muted reaction to a beat is consistent with AI skepticism – but the lock-up overhang muddies any clean read.
The 2027 commitment bill is the fulcrum: $22.2B is due either way. What changes is whether the revenue against it is diversified, durable, and large enough to outrun the depreciation now entering service.

Primary sources: SpaceX Form 10-Q for the quarter ended June 30, 2026 (filed 2026-08-04) and the same-day earnings call transcript. [10-Q] = filed report. [CALL] = unaudited management statements, not part of the filing. [3P] = third-party/press reporting, dated where known, including the PLTR valuation comp. The Customer A/B discussion is our inference (B from dated deal reporting, A from elimination reasoning), unconfirmed by SpaceX. The AI revenue table is a bridge of disclosed year-over-year changes; its residual line is the unexplained gap in that bridge, not a reported revenue line. Valuation figures are self-computed from cross-checked market cap and revenue data, not independently audited. Analysis of public disclosures, not investment advice, and not a recommendation to buy, sell, or hold any security.

Disclaimer: This article is for informational and educational purposes only and does not constitute financial, investment, tax, or legal advice. Any opinions, scenarios, price targets, or market observations reflect my personal views and may change without notice. Investing and trading involve substantial risk, including the possible loss of principal. You are solely responsible for your own investment decisions, position sizing, risk management, and trades. Conduct your own research and consult a qualified professional where appropriate.

DigitalOcean’s AI Pivot Is Accelerating. Can the Economics Keep Up?

Inder Sabharwal — Tue, 04 Aug 2026 17:22:34 GMT

What the company does and how it makes money

_{Source: digitalocean.com/products.}

DigitalOcean is trying to provide everything needed to build and run an AI application:
- It’s own GPUs and data centers, traditional cloud infrastructure and access to more than 70 AI models through one connection.
- Tools that give models access to private company data and managed AI agents.
- Because the entire stack operates within DigitalOcean customers can avoid some data-transfer fees.
- The economic bet is that owning the hardware will improve margins if demand keeps the GPUs busy.
- The risk is that DigitalOcean must spend the money before that demand and revenue arrive.
In production at named customers (company case studies): Character.ai serves 1B+ queries a day on its AMD Instinct GPUs; Workato runs 1T+ automation tasks on the Inference Engine, citing 67% lower cost.
No segment P&L is disclosed; ARR cohorts and an AI split stand in: AI Customer ARR $234M (+212% YoY), 20.8% of $1,125M total ARR (+29% YoY). 85% of AI ARR is from Inference Services and Core Cloud, not bare-metal rental (CEO); Inference grew roughly 800% YoY off a small base.
Large customers drive the mix: $1M+ ARR customers +214% YoY (23% of revenue); $500K+ 26% (+160%); $100K+ 35% (+98%). The top 25 are only 20% of ARR, so not concentrated, at least for now.

The business

Q2 revenue $281M, +29% YoY, accelerating from Q1’s +22% and beating expectations. Q3 guided to $304-307M (+32-34%); full-year raised to $1.17-1.18B (roughly 30.5% growth) with a stated Q4 exit rate of “35% or more.”
Adjusted profitability looks strong: 40% adjusted EBITDA margin ($114M, +27% YoY), 24% adjusted operating margin, and $63M of first-half adjusted free cash flow, in line with the 11-13% full-year guide.
GAAP tells a different story: operating income $29M (-18% YoY), net income $35M (-4% YoY), and first-half operating cash flow of $157M essentially flat (+0.2% YoY) against 25%+ revenue growth.
June 30 balance sheet: $767M cash against $921M total debt plus $578M of finance lease obligations. In July DigitalOcean equitized $472M of its 0% 2030 convertible notes with minimal cash and dilution; pro forma net leverage 0.7x.

The bull case

RPO (remaining performance obligations) of $894M, up 12x YoY, with $366M recognizable within 12 months, above the company’s own pre-announced “>$800M.”
A record $93M of incremental ARR (+191% YoY), the most in company history. Management said new data-center capacity contributed “very little” to it, pointing to demand against existing capacity.
The story itself is a market force: DOCN trades, in part, as a smaller, more direct claim on the AI-infrastructure buildout than a hyperscaler, without a neocloud’s commodity-rental shape. The gap is one customers can name: in a company-published testimonial, an AI-startup founder picked DigitalOcean because hyperscalers “don’t have any GPUs left” and cheap data centers have “fragile reliability,” while it offered “availability but also reliability at scale.” Naming it is not endorsing it: story-driven demand moves stocks independent of fundamentals, and it cuts both ways.

The bear case

GAAP operating income fell 18% and net income fell 4% while first-half operating cash flow stayed flat, even as revenue grew 25%+. Every profitability number celebrated on the call was adjusted. Neither management nor a single analyst addressed the GAAP decline. A working hypothesis, unconfirmed: the back-half-loaded buildout ($94M of first-half capex) is driving depreciation ahead of revenue.
The capacity build carries costs now, revenue later: the majority of the 155MW comes online by end of 2027, and the CFO suggested it is late in the year for new commitments to move 2027 much.
GPU unit economics are a black box: asked directly about AMD-vs-NVIDIA economics, the SVP of AI declined, a real non-answer on a material gross-margin driver.
No valuation read available. After a roughly +164% YTD run into the print (vs. Fastly +126%, Cloudflare +43%, Akamai +35%), the evidence speaks to growth quality, not to whether it is cheap or expensive.

What settles it

The Q3 2026 report: revenue vs. the $304-307M guide. At or above supports the acceleration story; a miss weakens the 35%+ exit-rate path.
A GAAP explanation. If buildout depreciation is the driver, disclosure should say so or the gap should narrow as capacity revenue lands; silence plus a widening gap weakens the adjusted-margin narrative.
Formal 2027 guidance, whenever it is issued. A number at or above 50% supports the reiterated framing; anything below weakens it.

Data and market conditions as of August 4, 2026. Historical snapshot, not auto-updated.

Premarket Report, Tuesday, August 4, 2026

Inder Sabharwal — Tue, 04 Aug 2026 13:24:38 GMT

Situational Awareness

🕊️ Iran talks status is contradicted between the two sides. Trump said Monday that talks with Iran were under way and he wanted to give Tehran “every last chance” before renewed strikes; Iran’s Foreign Ministry spokesman said the same day that Tehran has no meetings with US officials planned and is not engaged in talks. Markets are pricing Trump’s framing, not Iran’s denial, S&P 500 closed Monday just shy of its record (7,620.90), Nasdaq +2.1%, Dow hit a fresh record (53,178.41). Watch for the gap between the two sides’ public statements to matter if either escalates.
📡 The optics/photonics rally has a real, specific anchor: a reported FCC/Trump-administration plan to ban Chinese-made optical transceiver imports for AI data centers. Reuters reported the FCC is drafting a ban on next-generation Chinese transceiver modules, citing cybersecurity/data-leakage risk; China’s Zhongji Innolight (27% of global transceiver share) was already placed on the Pentagon’s military-backed company list in June. The policy would lock Chinese competitors out of a fast-growing market and funnel demand to US-based makers, direct beneficiaries are $AAOI, $COHR, $LITE, $GLW, $FN. This is a genuine regulatory catalyst, not generic sector sympathy off Monday’s Iran de-escalation, which is how this was first characterized here this morning before the specific driver was confirmed. The broader chip-equipment/memory basket (Marvell, SanDisk, Western Digital, Lam Research, and others still up 5-8% without their own fresh prints) is separate and does still read as Iran-de-escalation-driven sympathy.
🏗️ A mega-cap beat-and-raise morning: $PLTR and $CAT both posted major top-and-bottom-line beats with guidance raises. See Top Calls, this is today’s real signal, not the sympathy basket.
🔌 Second-degree AI read-through: $CAT’s Power & Energy segment (+17%, power generation specifically +29%) is explicitly tied to data-center buildout demand, a non-tech name capturing AI capex spend through power generation equipment, not chips.

Top Calls

$AAOI +17.4% (+20%+ intraday per some sourcing; up 60%+ from its recent lows) · regulatory · High, Direct beneficiary of the reported FCC plan to ban Chinese optical transceiver imports, a US-based maker of the exact hardware category the ban would protect. Not an earnings move (EPS/revenue surprise figures are stale prior-quarter data and don’t apply here).
$PLTR +15.3% · STRONG-borderline (70) · High, Adj EPS $0.41 vs $0.35 est (+18.5%), revenue $1.94B vs $1.80B est (+93% YoY), US commercial +149% YoY. Raised FY26 revenue guide from $7.65-7.66B to $8.15-8.16B, a real guide raise, not just a beat. 9th consecutive quarterly beat.
$CAT +11.5% · vol-gated SKIP, pending 9:30 confirmation · High, Adj EPS $8.17 vs $6.20 est (+31.8%), revenue $20.5B (+24% YoY, first quarter ever above $20B). Construction Industries +35%, Power & Energy +17% on data-center-tied power generation +29%. Raised FY26 revenue growth guide to mid-to-high-teens from low-double-digits.
$AMRC +39.8% · STRONG⭐ (75) · Medium, Continuing from yesterday’s after-hours print (revenue beat +10.6%, mixed EPS); premarket gap has extended sharply further this morning.
$W +18.9% · CAND+ (67) · Medium, Beat both lines by most accounts (EPS $0.95 vs $0.87 est, revenue $3.52B vs $3.45B est, +7.5% YoY), though sourcing disagrees on the exact surprise direction, see Setups.
AEIS * * + 14.8INSP +14.6% / LIFE * * + 23.1PAY +9.6% / $CTOS +10.1% · vol-gated SKIP, pending 9:30 confirmation · Medium, All five continuing from yesterday’s after-hours prints (see yesterday’s after-hours report for the original thesis on each); today’s premarket gaps are the cumulative move since yesterday’s regular close.
$GRAB +5.2% · news_driven, gap below scan threshold · Medium, Revenue $997M +22% YoY, adjusted EBITDA +54% YoY to $168M, net income +571% YoY. Raised full-year guidance, announced a $750M buyback. Below the 8% scan gate but a genuinely strong print from a name whose stock already moved most of this yesterday (+5.1% regular session, +3% AH), today’s premarket gap is a small residual.
⚠ $DOCN −9.6% · earnings, Q2 2026 · High significance, negative, Not on the premarket up-only scan (a documented screen limitation, it only filters gap-UP moves). Genuinely strong print (revenue $281M +29% YoY beat, RPO $894M up 12x YoY, guidance raised) selling off on a real GAAP profitability wrinkle. See Setups for the full read.

The Setups

$AAOI

Applied Optoelectronics is the clearest single beneficiary of a reported FCC/Trump-administration plan to ban Chinese-made optical transceiver imports for AI data centers (Reuters). The policy targets next-generation transceiver modules on cybersecurity/data-leakage grounds; China’s Zhongji Innolight, which holds roughly 27% of global transceiver share, was placed on the Pentagon’s military-backed company list back in June. AAOI makes the same category of hardware domestically and would be a direct share-gain candidate if the ban lands. Up 60%+ from its recent lows on this and the broader optics move; today’s premarket gap alone is 17%+. $COHR, $LITE, $GLW, and $FN are moving on the same catalyst, with AAOI showing the largest percentage response of the group.

$PLTR

Palantir’s Q2 2026 was a clean double beat with a guidance raise across every key metric management gives. Adjusted EPS $0.41 vs. $0.35 expected, revenue $1.94B vs. $1.80B expected, revenue growth accelerating to 93% YoY. US commercial revenue specifically grew 149% YoY to $764M, up 380% since 2024. Management raised full-year 2026 revenue guidance from $7.65-7.66B to $8.15-8.16B and US commercial guidance from $3.22B to “in excess of” $3.42B, a real forward raise, not just a backward-looking beat. This is the 9th consecutive quarter of beating estimates.

$CAT

Caterpillar’s Q2 2026 beat was unusually large: adjusted EPS $8.17 vs. $6.20 expected, a 32% surprise. Revenue $20.5B, +24% YoY and the first quarter the company has ever cleared $20B. Construction Industries led at +35% (North American construction demand +50%), and Power & Energy grew 17%, with the power-generation sub-segment, largely data-center-tied, up 29%. Management raised full-year revenue growth guidance to mid-to-high-teens from a prior low-double-digits target. The stock’s own relative strength has been weak coming in, so this reads as a reversal off underperformance more than confirmation of an already-strong trend, same shape as several other prints this cycle.

$AMRC

Continuing from last night’s after-hours print (revenue beat 10.6%, mixed EPS). The premarket gap has extended sharply since, from roughly 24% after hours last night to nearly 40% this morning, with the conference call’s color from yesterday now more fully priced in.

$W

Wayfair’s Q2 2026 beat both lines by most reporting: EPS $0.95 vs. $0.87 expected, revenue $3.52B vs. $3.45B expected (+7.5% YoY). One flag: sourcing disagreed on the exact EPS surprise direction and magnitude, and on the immediate market reaction, this wasn’t resolved before publish, so treat the specific beat percentage as approximate pending a cleaner read later today, not the headline beat-and-raise itself.

$AEIS / $INSP / $LIFE / $PAY / $CTOS

All five are continuing from yesterday’s after-hours prints, see yesterday’s after-hours report for each name’s original thesis (AEIS: FY26 growth target raised, data-center revenue +102% YoY structural story; INSP: EPS surprise +159% following a guidance-cut quarter; LIFE: large revenue beat, small-cap with elevated short interest; PAY: clean beat backed by strong existing trend; CTOS: EPS beat, call happened this morning). Today’s premarket gaps are each name’s cumulative move since yesterday’s regular-session close, not a new catalyst.

$GRAB

Grab’s Q2 2026 print was strong across the board: revenue $997M (+22% YoY), adjusted EBITDA +54% YoY to $168M, more than double the pace of revenue growth, meaning real operating leverage, not just top-line growth. Net income +571% YoY to $235M. On-Demand GMV +21% YoY to $6.5B with a record 54M monthly transacting users. Management raised full-year guidance and announced a $750M share buyback. Most of the move already happened yesterday (+5.1% regular session, +3% after hours); today’s premarket gap is a smaller residual on top of that.

$DOCN

DigitalOcean’s actual Q2 2026 numbers were strong: revenue $281M (+29% YoY, beat), RPO $894M (up 12x YoY, beat the company’s own pre-announced “>$800M” guide), AI Customer ARR +212% YoY to $234M, record $93M incremental ARR (+191% YoY). Full-year guidance was raised to $1.170-1.180B (30-31% growth), with Q3 guided even higher at 32-34% YoY. Despite all of that, the stock is down roughly 9.6% this morning. The specific wrinkle: operating income fell 18% YoY and net income fell 4% YoY even as adjusted EBITDA margin held at 40%, a real GAAP-vs-adjusted profitability gap inside an otherwise excellent print, combined with a stock that had already run ~150% YTD and pre-announced most of this good news on 7/7. Reads as sell-the-news on a real margin wrinkle, not a demand problem.

R10.1 Second-Entry Watch

$KNSA (STRONG⭐⭐ on 7/28) is at day 5 of its second-entry window today, the tail end of the days-4-6 range this framework watches.

Bottom line

Today’s highest-conviction print is $PLTR, a real guidance raise on top of a double beat, not just a number that cleared a bar.

📊 Read this on SignalDeck.live.

Not financial advice. For informational and educational purposes only.

$PLTR: a sovereign-AI arms dealer printing 93% growth, with America driving ~90% of the new revenue

Inder Sabharwal — Tue, 04 Aug 2026 12:39:29 GMT

What The Company Does And How It Makes Money

Palantir sells operating-system software for institutions: Gotham for defense and intelligence, Foundry for corporations, and AIP (its Artificial Intelligence Platform, the layer that puts LLMs to work on a customer’s own data) across both.
US government: $809M, 42% of revenue, +90% Y/Y. The original franchise, exposed to US government procurement and budget cycles, still growing rapidly.
US commercial: $764M, 39% of revenue, +149% Y/Y. The line most closely tied to the AIP and sovereignty thesis, and the fastest-growing of the three.
International (commercial + government): ~$362M, 19% of revenue, +33% Y/Y (derived from filed segment math). Company-wide adjusted gross margin is 86%; per-line margins are not disclosed.

The Business

Revenue $1.935B, +93% Y/Y, +19% Q/Q, against a raised full-year guide of $8.150-8.158B (+82%).
Margins expanded while growth accelerated: 47% GAAP operating margin, 62% adjusted; Rule of 40 (growth rate plus adjusted operating margin) of 155. This quarter’s GAAP net income ($1.062B, 55% margin) exceeded ALL of Q2 2025 revenue ($1.004B), arithmetic verified against the filing.

Cash machine: $1.22B Q2 adjusted free cash flow (63% margin); $9.2B in cash and Treasuries with zero debt as of June 30, 2026.
Q2 2026 billings ($2.072B) exceeded recognized revenue in dollars, +88% Y/Y; RPO (contracted, non-cancelable future revenue) $4.90B, +102% Y/Y; net dollar retention 157%.

The Bull Case

The growth-profitability combination is the story: 93% growth, 62% adjusted operating margin, GAAP profitable, against an ~$8.15B FY revenue guide.
Contract-value indicators remain strong: record $2.132B US commercial TCV (total potential lifetime contract value closed), +153%; US commercial RDV (remaining deal value) $6.238B, +124%; US commercial customers +35% Y/Y to 653 (TTM count, June 2026).
Everything raised mid-year: revenue, US commercial revenue (at least 134% growth), adjusted operating income, and adjusted FCF ($4.5-4.7B), extending an eight-quarter beat-and-raise pattern through Q2 2026.
Management’s thesis, now in its own words: customers “have declined to become vassal states of the language labs,” and Palantir is paid, in Karp’s words, “as a derivative of value creation,” rather than “for clicks or tokens or chats” (shareholder letter). Management’s framing, not established fact, but the bookings are consistent with it.

The Bear Case

International produced under 10% of the growth. It was 27% of prior-year revenue but contributed just $90M of the $932M Y/Y increase, growing 33% against the 93% headline. Mix fell from 27.1% to 18.7%; the “global platform” is a US-demand story.
~44x forward sales, after the raise. Using diluted weighted-average shares (2.569B) as a rough proxy for period-end shares, the $139.44 AH print implies roughly $358B of diluted value against the $8.154B guide midpoint, a directional gauge, not an official market-cap print, with no view here on whether it is justified.
Guidance itself encodes deceleration. The Q3 guide ($2.160-2.164B) plus the full-year midpoint imply Q4 of ~$2.42B, roughly +72% Y/Y, down from 93% now.
The bookings assume nothing walks. By the company’s own definitions, TCV and RDV presume every contract option is exercised, and most contracts allow termination for convenience (their own risk disclosure).

What Settles It

The Q2 10-Q and any official call transcript: the RPO step-change (Q3’25 $2.60B to Q4’25 $4.08B) remains unexplained in the same-day reporting reviewed; an explanation, plus international pipeline color, supports or weakens the durability read.
The Q3 print (around early November 2026, typical cadence): international growth reaccelerating toward the company average supports the global story; another quarter of ~30% Y/Y growth weakens it.
Q3 actual vs. the $2.16B guide. A ninth straight beat-and-raise supports the bulls; an in-range landing supports the guided glide toward ~72%. Karp set his own bar on the call, growth “equal to or above what we have in U.S. commercial for the next 18 months” (secondary reporting; a stated aspiration, not formal guidance). The next two prints provide the first tests of it.

Bottom Line

The reported quarter is exceptionally strong. Growth, margins, and cash all strengthened; the softer spots (international mix, total TCV +49% lagging revenue +93%, a cautionary signal though TCV includes optional and terminable value, so not a clean apples-to-apples) live in the bear case, not the headline numbers.
You are underwriting something narrower than the headline: growth is disproportionately US commercial and US government; international is real but growing far slower.
The multiple prices continuation while the guide points down: at ~44x forward sales, the market is paying for the beat-and-raise pattern to persist through a guided deceleration to ~72%. Even after the +10.8% AH move ($125.86 to $139.44), the stock sits ~33% below its $207.52 52-week high.
Watch for two missing facts before extrapolating: an RPO step-change explanation (management may never provide one) and any sign of international reacceleration over the next quarter or two.

The Bull Case for GitLab $GTLB

Inder Sabharwal — Sun, 02 Aug 2026 20:32:21 GMT

Deep-dive, July 28, 2026. GitLab (NASDAQ: $GTLB) reported its latest quarter on June 2, 2026. Price data as of July 28, 2026. A snapshot in time, not auto-updated. Educational only, not financial advice, we never predict where a stock is headed.

📊 Read this on SignalDeck.live.

What GitLab does

One workbench for the entire software-building process: writing code, running automated tests, scanning for security bugs, and tracking who’s working on what, all in one system instead of stitching together separate tools from different vendors.
The customer pitch: one platform, one audit trail. Engineering teams get faster releases and a single place to see everything that happened to their code, which matters most to bigger, regulated companies.
Customers span scrappy startups to regulated banks and large enterprises (Zillow and CSL Behring are named customers this quarter), running GitLab either on their own servers or via GitLab’s cloud.
Where AI actually fits: the coding and review work itself. GitLab’s AI agent product (called “Duo”) can write, review, and test code alongside, or increasingly instead of, human engineers. Three ways AI shows up in the bull case below: new usage-based AI revenue, a new toll-booth product that gets paid even when a customer uses a rival’s AI coding tool, and GitLab’s governance and audit tools getting more valuable, not less, as AI agents write more of the actual code.

What this actually is

A company that just grew revenue 23% while guiding to only 16-17% for the year, at about $34.66 a share and a $5.85B market cap. The bull read: the stock price seems to reflect the cautious guidance, not what the company just delivered, plus a growing pile of near-term signed business, the fastest new-customer growth in 10 quarters, and an AI product management isn’t even counting in its own numbers yet.
The setup: the core, boring business carries the case on its own. Any real AI revenue on top is pure upside to numbers the company just beat.

The boring, durable stuff is working

Revenue $264.2M, up 23% year-over-year, four points ahead of guidance, against a full-year guide of just 16-17%.
Signed-but-not-yet-billed work due in the next 12 months grew 24%, faster than the total backlog (18%), a good sign that more of what’s sold is set to show up as revenue soon rather than sit on the books.
Existing customers spend 17% more per year than they did a year ago, and over 90% stick around at all, growth GitLab gets almost automatically before signing a single new customer.
Customers paying over $100,000 a year grew 18% and now make up more than three-quarters of all revenue, the push upmarket is working.
New customer signings hit their highest count in 10 quarters, up 30% year-over-year, a leading signal for future growth since new customers take time to grow into big accounts.
Adjusted profit margin hit 14%, up about 2 points from a year ago, alongside $146.7M in free cash flow (a 56% margin, helped some this quarter by faster bill collection, so not fully repeatable, but still a strong quarter).
$1.36B sitting in cash and short-term investments, plus roughly 2.4 million shares bought back this quarter and $350M left to keep buying, a real capital-return signal.
GitLab is cutting costs hard: about 14% of its workforce and operations in 22 countries, restructuring charges of $30-35M. Savings get reinvested into the AI push below, and profitability is expected to bottom out this quarter before improving.

The AI upside nobody’s paying for yet

GitLab’s AI agent product (DAP) launched two weeks before the quarter and already generated more new revenue in its first quarter than its two older AI products combined ever did in any single quarter.
It’s already being paid for on a usage basis, close to $20 million a year run-rate at quarter-end. GitLab’s own finance chief told analysts not to build models around that number yet, since it’s just one quarter of data, a refreshingly honest caveat, not a red flag.
A top-10 US bank piloted it and saved 1.5 hours per coding task, with plans to expand the user base roughly 20x.
It’s now sold through Amazon, Google Cloud, and Anthropic’s marketplaces, making it much easier for big companies to actually buy.
Management’s own full-year guidance assumes this AI product contributes basically nothing material. Everything above is upside sitting on top of numbers the company already just beat.
GitLab is also building bigger bets: a rebuilt version of Git (the code-tracking technology nearly every developer uses) done in partnership with an unnamed AI lab, aiming for 100x more scale, and a new product called “GitLab Orbit” that other AI coding tools, not just GitLab’s own, can pay to plug into.

Why this might be a moat, not just a feature

One platform for every way software gets built: by hand, AI-assisted, or fully autonomous, while most competitors only cover one of those. As more coding shifts to AI agents, having one company handle security, audit trails, and permissions across all of it arguably gets more valuable, not less.
Win rates against GitHub (its biggest rival, owned by Microsoft) improved slightly this year, while new-customer signings hit a 10-quarter high.

What the stock costs

About 4 times next year’s expected sales once you subtract GitLab’s cash pile, a reasonable price for a company growing in the high teens to twenties percent with improving margins and an AI product line that isn’t even being counted yet.
Not a screaming bargain on every measure: the company is still posting a loss under standard accounting rules, and its forward earnings multiple (42.6x) isn’t cheap.

The honest counter-case

Overall bookings growth slowed sharply, to 12% from 35% a year ago, a tough comparison, but a real slowdown nonetheless.
Revenue from customers running GitLab on their own servers was flat for the first time since the company went public. Cloud growth is masking that in the blended number.
Smaller customers (paying $5,000+ a year) are barely growing, up just 7%, versus 18% for the big accounts.
About a fifth of GitLab’s revenue comes from price-sensitive customers the company doesn’t expect to recover this year.
Layoffs at GitLab’s own customers, plus one-time M&A-related churn, cost GitLab more seats than expected this quarter.
Companies are getting more hesitant to sign multi-year contracts given the uncertain economy.
That AI revenue number is one quarter old, and the company’s own CFO says not to extrapolate it.
The cost-cutting carries real execution risk: 14% of staff and 22 countries exited, while trying to pull off a major technical rebuild at the same time.

Bottom line

The stock’s price seems to reflect GitLab’s cautious guidance (16-17% growth), not what it actually just delivered (23%) or the improving signs underneath it.
The AI product is early-stage but already real, and it isn’t even counted in the numbers yet, that’s the asymmetry bulls are underwriting.
The single most interesting data point: an AI lab chose to build a rebuilt version of Git with GitLab rather than work around it, a vote of confidence in GitLab’s relevance from inside the AI industry itself.
This only works if GitLab keeps winning enterprise consolidation, its usage-based pricing (“Flex”) lands, and a leaner organization pulls off a technical overhaul without dropping the ball. Watch flat on-prem revenue, weak small-customer growth, and slowing bookings for signs it isn’t.
Read this as a growth-plus-upside story, not a deep-value one: the accounting loss and 42.6x earnings multiple mean the core business has to keep performing for the AI upside to matter.

Educational content, not financial advice. For informational purposes only. Do your own research.