TMT Breakout晨报汇总多只科技股财报与盘前/盘中反应

Good morning. **QQQs +20bps (SPX +40bps)** as stocks look to continued recent rally. Overall, not much on the macro front overnight.

In Tech, **Semis -1%** and **Software +1.5%** as **AMD -9%** disappointed on NT numbers while Lisa did her best to keep investors focused on ‘27 and beyond. **SPCX** **-11%** weaker on higher capex and ahead of the lock-up; Elon did a better job for **NVDA +1.5%** and **META +2%** on the call, raising ‘27 GW ambitions to 10GW and saying they are an exclusive **NVDA** shop, while saying compute payback is 1 year, which has investors wondering why **META** doesn’t turn on that rev spigot immediately. **SHOP +22%** and **ANET +11%** crushed and guided above, **BKNG +9%** better than feared, **ALAB -1%** crushed but LT questions remain.

We’ll cover **SPCX AMD ANET ALAB BKNG** and **SHOP** below then onto the usual.

Another busy one after the close today:


Let’s get to it…


### **SPCX: Big AI-led beat and pulled the revenue curve forward by roughly a year as Elon talked up compute ROI but higher capex and upcoming lock up weighing on the stock **

Musk with interesting datapoints during this call. He said SPCX is an NVDA exclusive shop as he thinks Rubin Ultra is the best machine — no ASIC build = higher priority from Jensen. He also said he is ramping FY27 capacity to 10GW, implying 5+GW of new capacity (street at 3.7GW), implying something on the order of ~$300B for AI Capex, depending on what $B/GW assumption you use, and an open question remains where that $ will come from.

Here he is on compute payback:
**

*"The current economics have translated into a** less than one year payback on our new capital deployments for compute.** For example, in the first few weeks of the third quarter, we've already contracted an additional $6.7 billion of cloud services revenue over a six month period that begins ramping starting in October of this year. We believe this puts us on a trajectory including contribution from Cursor to reach 100 billion of RR, or annualized revenue run rate, by the end of this year, based on our expected revenue in the month of December of this year"*

One would think this would be good for **META’s** compute ambitions…

On Memory:
**

*“Limiting factor currently is memory. Memory output is increasing by around 20% per year. Now normally that would be fantastically fast and amazing for any large mature industry. But ask yourself is the demand increasing by 20% a year? No, the demand is increasing by 200% a year, maybe higher. So if you’ve got demand increasing much faster than supply then Economics 101 would suggest that the price increases. It does not decrease.”*

**SPCX IPO Lock-Up Expirations**

**Date | Shares Unlocked | Value | % of SpaceX Market Cap**
8/6 | 911.5M | $123.1B | 7%
8/20 | 319.0M | $43.1B | 2.5%
9/9 | 319.0M | $43.1B | 2.5%
9/10 | 59.1M | $8.0B | 0.5%
9/24 | 328.4M | $44.3B | 2.5%

### **The #s:**

**Revenue** $7.814B, +91.9% y/y (last q +15.4%) vs Street $6.819B, +~67.5%; Adj. EBITDA $3.538B, +191% y/y, 45.3% margin vs Street $2.046B, 30.0%.
**AI** revenue $2.561B, +247% y/y and +213% q/q vs Street $2.016B, while AI Adj. EBITDA inflected to $1.146B vs Street $65M
**Q2 capex** was $18.4B, including $15.8B for AI compute, and mgmt said Q3 and Q4 should be “very similar.” vs Street at roughly $13.0B and $13.7B, respectively.

**Outlook** Mgmt expects more than $100B of total-company annualized December revenue, over 2GW of compute by YE26 and closer to 10GW than 5GW by YE27, but Q3 and Q4 capex should each be similar to Q2’s $18.4B.

### **Key Takeaways:**

-

**Run-rate** SPCX signed another $6.7B six-month cloud-services agreement that begins ramping in October, and mgmt expects this deal, existing contracts, Connectivity growth and Cursor to put total-company annualized revenue above $100B in December. This is not traditional software ARR; it is December monthly revenue annualized. Mgmt characterized $100B as effectively a floor, and the sell-side translation is low-$20Bs of Q4 revenue vs pre-print Street at $16.5B.

-

**Compute** Nameplate compute reached 1.4GW from 1.0GW last quarter and 0.4GW a year ago. Mgmt expects more than 2GW by YE26 and “closer to 10GW than 5GW” by YE27, well above pre-print assumptions around 4.2GW. SPCX is targeting a 20GW power, cooling and electrical project funnel, while acknowledging that roughly 15GW may actually be available if one-quarter of projects slip.

-

**Returns** Mgmt’s defense is that new compute deployments currently have a payback of less than one year, making AI capex closer to rapidly monetized capacity than conventional long-duration infrastructure. Q2 FCF of roughly -$14.9B was actually better than Stree

-

**Grok** Grok 4.5 token consumption tripled immediately after launch, Grok 4.6 is expected shortly, Grok 4.7 three to four weeks later and Grok 5 by year-end. Grok 5 will incorporate SPCX’s full data corpus, while mgmt expects only roughly 10% of total compute to be reserved for Grok training over time, leaving the majority available for third-party training and inference. Cursor is expected to close soon and is included in the December run-rate framework.

-

**Semis** SPCX is standardizing exclusively on NVIDIA Vera Rubin, expects a “very significant percentage” of NVIDIA’s 2027 GPU supply and plans to use an optimized NVL72 design both terrestrially and in Starmind orbital systems. Mgmt also argued that memory output is growing roughly 20% annually against demand of 200% or more, implying continued tightness and higher rather than lower pricing.

-

**Starlink** 1.7M net adds vs 1.4M last quarter, 12.0M subscribers and ARPU stable sequentially at $66. Most of the upside came from enterprise and government, where revenue grew 108% y/y, SPCX won more than $6B of U.S. contracts and aviation remains less than 10% penetrated. Mgmt said it has never lost an enterprise customer and ultimately expects enterprise revenue to substantially exceed consumer revenue

-

**Mobile** Starlink Mobile could become another enormous TAM, but also another source of capex uncertainty. Mgmt framed the next-generation system as roughly 100 times better, combining 65MHz of EchoStar spectrum vs 5MHz today with roughly 10 times more satellites. SPCX also confirmed that it intends to build a U.S. terrestrial layer, potentially using small base stations attached to existing Starlink installations, but declined to quantify the investment.

-

**V3** Mgmt’s Starlink thesis rests on a nonlinear capacity step-up: V3 satellites are expected to offer more than 10 times V2 capability, and SPCX expects to launch roughly 10 times as many. However, service will not inflect immediately; mgmt said it needs approximately 1,000 V3 satellites, likely around Q2’27, before customers see a meaningful improvement

-

**Starship** Flight 13 materially improved technical confidence. Mgmt said the heat shield appears robust and later described the heat-shield problem as “solved,” with no visible technical obstacles to full and rapid reusability. Flight 14, tentatively at the end of August, could attempt a tower catch and will be the first operational deployment of V3 broadband satellites. The one-flight-per-day target within roughly a year remains highly aspirational, but a successful catch is the next major non-AI catalyst.

### **Bull vs. Bear Debate**

**The bull case** starts with the idea that AI infrastructure has moved from speculative optionality to an operating business with contracted demand and exceptional incremental economics. SPCX generated $1.6B of incremental AI-infrastructure revenue in Q2, turned AI Adj. EBITDA positive at $1.1B, signed another $6.7B six-month agreement and claims current compute capex pays back in less than one year. If SPCX can repeatedly convert power and GPUs into contracted revenue faster and more cheaply than conventional data-center operators, the current spending is not a cash sink. It is the inventory build for an unusually high-return manufacturing and infrastructure model. Exclusive access to Rubin, a significant share of NVIDIA’s 2027 supply and a 15GW to 20GW power funnel would reinforce the time-to-power advantage.

The broader bull case is that AI, Starlink and Starship reinforce one another rather than operating as separate businesses. Starlink provides recurring cash generation and a growing enterprise and government franchise. Starship lowers launch costs, enables much larger V3 and Mobile constellations and eventually makes orbital compute feasible. AI and robotics then increase bandwidth demand, while SPCX’s launch and satellite capabilities let it deliver the network at structurally lower cost. This quarter strengthened that flywheel: Starlink added a record 1.7M subscribers, enterprise and government grew 108%, Flight 13 improved confidence in reusability, and V3 deployment is moving from concept toward an operational launch.

**The bear case** is that Q2’s AI upside came from short-duration capacity-offload agreements priced during an acute compute shortage, not from a durable cloud or software franchise. Some contracts preserve customer cancellation options after an initial period, and mgmt’s $30 to $50 per-watt monetization framework may normalize as Rubin supply, power availability and competing capacity expand. Even a modest decline in price per watt matters enormously when SPCX is spending roughly $18B per quarter today and potentially well over $100B annually in 2027. The less-than-one-year payback is a point-in-time claim; if utilization, contract renewals or pricing disappoint, the same capex can quickly become a large stranded-cost problem.

The rest of the portfolio still requires major execution. Starship’s heat shield may look solved, but full reuse, daily cadence, regulatory approvals and dependable payload deployment remain unproven. V3 does not materially improve Starlink service until roughly 1,000 satellites are deployed, which pushes the key capacity inflection into 2027. Consumer Starlink KPIs were mostly in line, blended ARPU is already down 22% y/y, and the quarter’s Connectivity upside was concentrated in government and enterprise contracts. Mobile adds another very large TAM, but SPCX has not quantified the terrestrial-network investment. Add the near-term lockup supply, and bears can argue that the operating story improved without reducing the stock’s financing, execution or duration risk.

### **AMD -9%: AMD delivered a clean beat/raise vs. streetand materially raised the 2027 Data Center earnings framework and blessed buyside ‘27/’28 #s, but near-term guide missed an elevated buyside bar, particularly on GM and the implied AI GPU ramp.**

Stock leaned long into the print and had outperformed semis leading up to their AI day/Anthropic announcement, but #s missed high expectations going in. Lisa tried to keep the focus on Q4/’27 and beyond ramp, essentially seemingly blessing higher than street numbers, but bulls will be disappointed that NT #s weren’t higher.

Buyside numbers likely unchanged after yesterday for FY27/FY28 with most sitting in low 20s/$30, respectively. We don’t think the print does anything to change the medium/long term CPU + GPU bull case, but unclear if this going into the funding short camp NT given lack of catalysts until we get closer to next earnings as Helios inflection begins at end of Q3. It wouldn’t surprise us if somewhat range bound, unless the whole group breaks out with low end at low $20s x 20 = $400-$440. Bulls will keep their LT framing as 25x+ on FY28 #s ($30) = ~$750. Risk/reward still seems pretty solid here.

### **The #s:**

**Q2** Revenue was $11.536B, +50.1% y/y (last q +37.8% y/y) vs Bogeys closer to $12B and Street $11.309B, +~47.2%; non-GAAP EPS was $1.66 vs Street $1.62.

Data Center revenue was $6.718B, +107.3% y/y vs Street $6.554B, +~102.3%, with server CPU revenue up >70% and estimated AI GPU revenue of ~$2.8B vs Street ~$2.65B

**FQ3’26** revenue guidance was $13.0B at the midpoint, +40.6% y/y vs Bogeys closer tto $13.2B and Street $12.502B, +~35.2%; implied EPS was ~$1.93 vs Street $1.87, while GM of 56.0% was 20 bps below Street

### **Key Takeaways**

-

**2027 DC** Mgmt’s explicitly called for server CPU revenue to grow >70% y/y in 2027 and total Data Center revenue to grow “well over” 100%. That compares with pre-print Street assumptions of roughly 52% server growth and 88% Data Center growth.

-

**Server CPU** Server is emerging as a second major AI growth engine alongside Instinct. Cloud and Enterprise server revenue each grew >70% y/y, with double-digit unit and ASP growth, skewed more toward units. Mgmt now expects >80% server growth in 2H26 and >70% in 2027 as agentic AI expands the CPU workload opportunity and Venice begins ramping. Customer demand for Venice is stronger than for any prior EPYC generation.

-

**AI GPU** Estimated FQ2 AI GPU revenue of ~$2.8B exceeded Street at ~$2.65B, but FQ3 implied AI GPU revenue of roughly $3.4B was only around Street at ~$3.36B. That likely explains some of the disappointment despite the headline revenue guide. The more meaningful MI450/Helios inflection begins at the end of FQ3, becomes much larger in FQ4 and steps up again in FQ1’27.

-

**Helios** Customer pull is tracking ahead of AMD’s original forecast, specifically in terms of anticipated 2027 volumes rather than current yields. OpenAI, Meta and Anthropic are the strategic anchor customers; Microsoft will deploy Helios at scale on Azure, and mgmt said there are additional sub-gigawatt customers. The product is in production, initial shipments begin late FQ3, and yields should improve over the first several quarters of the ramp.

-

**Supply** Mgmt believes AMD has enough supply to exceed the stated guidance and sees upside potential, with strong visibility into 2027 HBM allocation. However, the current server supply chain remains tight because much of the demand was unforecasted. FQ2 capex jumped to $808M from $389M as AMD invested in backend and packaging equipment

-

**Margins** FQ2 GM of 56.2% was slightly above Street at 56.1%, but FQ3 GM guidance of 56.0% was below Street at 56.2%, and Opex guidance of $3.65B exceeded Street at $3.535B. Server CPUs and the Embedded recovery are margin-accretive, while Data Center AI GM remains slightly below the corporate average. The callback indicated Opex could grow ~35% in 2026 and at a similar rate in 2027, although revenue estimates are also rising substantially.

-

**Embedded** Embedded revenue was $977M, +18.6% y/y vs Street $960M, +~16.5%, marking the strongest growth in more than three years. Demand improved across networking, aerospace and defense, test and measurement, and communications; the segment carried a 40% operating margin and AMD is tracking toward >$18B of new design wins. FQ3 Embedded revenue is expected to grow strong double digits q/q and was one of the primary sources of guide upside.

-

**Long Term** AMD raised the overall high-performance and AI-compute opportunity to nearly $2T by 2030, growing approximately 40% annually. Mgmt expects AMD revenue to grow above that market rate, substantially ahead of its prior >35% target, and to significantly exceed its prior $20 annual EPS target within the original strategic timeframe

-

**Callback **The most important callback-specific incremental item was Opex: AMD indicated it will likely rise approximately 35% in 2026 and by a similar rate in 2027, materially above prior models, although the revenue outlook has also moved substantially higher. The post-call analyst meeting otherwise reinforced the key call messages: server revenue >80% y/y in 2H26 and >70% in 2027, Data Center revenue more than doubling in 2027, a softer 2H PC market due to memory and component costs, and initial Helios shipments late FQ3 followed by the larger FQ4/2027 ramp.

### **Bull vs. Bear Debate**

**The bull case** is that AMD is becoming one of the few companies with two independent but mutually reinforcing AI growth engines. Instinct gives AMD exposure to the rapidly expanding accelerator market, while EPYC is benefiting from both traditional server share gains and the emergence of agentic AI workloads. FQ2 provided strong evidence for the CPU thesis: Cloud and Enterprise revenue each grew >70%, both units and ASPs rose double digits, and the growth was more unit-driven, indicating real volume and share gains rather than simply richer mix. Venice demand is stronger than for any previous EPYC generation, and the 2027 server growth outlook of >70% is particularly notable because it comes off a much larger 2026 base.

The accelerator case has also become more tangible. Helios is now in production, expected demand is ahead of AMD’s initial forecast, and the platform has several large anchor customers plus Microsoft and additional smaller deployments. The ramp is not a single-quarter event: FQ3 is the beginning, FQ4 is a meaningful step-up, FQ1’27 rises again, and AMD expects continued growth through 2027. ROCm now supports more than three million models out of the box, open-source contributions rose more than tenfold, and AMD is co-optimizing software with the leading AI labs. Bulls argue that this reduces the historical software disadvantage while the annual rack-scale cadence through MI500 gives customers confidence that AMD can remain a durable second merchant platform rather than a one-generation alternative.

What Changed This Quarter : The bull case became materially stronger on server CPUs, 2027 visibility and supply confidence. Before the quarter, investors primarily viewed the next leg of the AMD story through Helios. The call established that EPYC is also accelerating because agentic AI is expanding the CPU TAM, and mgmt was willing to guide to >70% server growth and well over 100% total Data Center growth with five months still remaining before 2027. Helios demand being ahead of AMD’s original volume assumptions, broader customer diversity and the return of Embedded growth added further support.

**The bear case** begins with the fact that near-term results were only modestly better than expectations despite an extremely favorable AI demand environment and a stock that had already risen more than 130% YTD. FQ2 revenue and EPS beat Street but missed bogeys, but FQ3 GM was slightly below Street, Opex was meaningfully above, and implied FQ3 AI GPU revenue was approximately in line. Bears argue that the market is no longer paying for a normal beat and raise; it is already underwriting a large portion of the 2027 Helios and server ramp, so anything short of major estimate increases can drive multiple compression.

Execution risk is also rising as AMD moves from accelerator cards into a full rack-scale platform. Helios requires AMD to coordinate CPUs, GPUs, networking, software, HBM, advanced packaging and customer data-center readiness. Mgmt is confident, but the current server supply chain is tight and capex has risen sharply to secure backend capacity. Some estimate that meeting AMD’s 2027 targets could require materially more CoWoS capacity than is presently assumed, with packaging yields and substrate availability additional risks. Data Center AI is also below corporate GM, and Opex may grow ~35% in both 2026 and 2027. Bears therefore question how much operating leverage will actually be realized if the fastest-growing revenue stream is margin-dilutive and requires heavier-than-expected R&D and supply-chain investment.

The other major debate is the quality and durability of the AI GPU revenue. OpenAI, Meta and Anthropic provide enormous volume, but customer concentration is high and several deployments are supported by warrant arrangements. Those warrants can create substantial economic dilution that is not fully reflected in headline non-GAAP EPS. Competition also remains intense from NVDA, custom accelerators, AVGO/MRVL silicon programs and eventually improved INTC and ARM server products. AMD still needs to prove that Helios can deliver competitive performance, software usability and customer ROI at scale, not just win strategically motivated second-source commitments.

### **ANET +11%: delivered a monster beat-and-raise above bogeys: Q2 revenue, EPS and margins cleared Street, Q3 growth accelerates to 43%, and FY26 growth guidance jumped to 40% as improved supply finally unlocked broad-based demand.**

Expectations were high going in for one of the few AI stocks hitting 52wk highs, but revs came through here, well above any of the buyside bogeys we heard going in and 3rd straight q of material acceleration. Buyside EPS likely moving close to $7 after this print.

### **The #s:**

**Q2 Revenue** was $3.036B, +37.7% y/y (last q +35.1%) vs Street $2.830B, +~28.4%.
EPS was $1.02, +39.7% y/y vs Street $0.89; non-GAAP GM/OM were 63.4%/49.9% vs Street 62.2%/46.5%.

**Q3 revenue guidance** was ~$3.3B, +43.0% y/y vs Street $2.951B, +~27.8%, with EPS of $1.06–1.08 vs Street $0.92.

**FY26 revenue** **guidance** rose to ~$12.6B, +40% y/y vs Street $11.617B, +~29% and buyside expectations around +35%; the non-GAAP OM target increased to 48–49% from 46%

### **Key Takeaways:**

-

**Guide** The $1.1B FY26 revenue raise was far larger than investors expected and was the clear stock-moving item. Q3 revenue guidance was nearly 12% above Street, FY26 guidance was ~8.5% above Street and meaningfully above the roughly 35% growth buyside bogey, while the OM target moved up 250bps at the midpoint. The quarter effectively reversed the Q1 narrative that supply constraints would cap ANET’s ability to convert strong demand into revenue.

-

**Deferred** Total deferred revenue rose to $6.9B, +69% y/y and +11% q/q, while product deferred increased roughly $600M sequentially despite the large revenue beat. **The main bear nit** is that billings growth slowed to 16% y/y from 54% last quarter, although deferred continued to rise and mgmt repeatedly cautioned that acceptance clauses make the metric volatile and disconnected from underlying demand.

-

**Demand** The upside was much broader than just AI back-end networking. Mgmt called out core data-center front-end, enterprise, campus, routing and AI fabrics, and explicitly said it believes both the AI and campus numbers will increase even though the formal targets were left unchanged. The incremental $1.1B is therefore not a single-customer or single-product event; mgmt expects contributions across every major product category.

-

**Supply** This was the biggest narrative change. ANET secured 2026 memory supply, extended visibility well into 2027 across DDR4, DDR5 and NAND, added suppliers, improved delivery terms, increased factory and distribution capacity, and established a liquid-cooling supply chain. Purchase commitments reached $9.7B, nearly tripling y/y. Supply is still tight industry-wide and mgmt does not expect normalization until 2028, but ANET now appears to have secured a larger allocation of what is available.

-

**AI** Etherlink AI fabrics now have more than 100 cumulative customers versus four to five in 2024. Scale-across revenue is expected to be roughly $1.2B in 2026 within a current $3B–$4B market that mgmt believes can grow to $15B–$20B by 2030. Scale-out and scale-across remain the immediate opportunities, while scale-up becomes more meaningful as non-NVIDIA accelerators, custom XPUs and open Ethernet alternatives gain adoption.

-

**Margins** Q2 GM of 63.4% beat Street by roughly 120bps, but only 20–30bps of the upside versus guide came from tariff refunds; most of the remainder came from customer mix. Price increases have not materially entered reported results and should begin helping toward year-end and into 2027, providing an offset to higher memory and silicon costs. Operating leverage was exceptional, with OM at 49.9% and the FY26 target raised to 48–49% from 46%.

-

**Competition** Mgmt framed white box as a viable tactical option for simpler or highly staffed deployments, but argued that network complexity, reliability, security, job-completion time and operating costs increasingly favor the integrated EOS/hardware approach. The main competitive limitation is scale-up inside NVIDIA’s vertically integrated NVLink stack, where ANET currently has little participation. Its opportunity is stronger in scale-out, scale-across and non-NVIDIA/custom-accelerator architectures.

-

**Customers** Microsoft and Meta are expected to remain 10% customers, and mgmt continues to expect one, potentially two, additional 10% customers. ANET did not identify them or commit to timing, but improved supply makes it more feasible for those ramps to reach recognized revenue during 2H26.

### **Bull vs. Bear Debate**

**The structural bull case** is that ANET is becoming the networking control point for the AI infrastructure build, not simply a vendor of back-end Ethernet switches. AI clusters need higher utilization, lower latency, deterministic routing, security, rapid software upgrades and connectivity between geographically distributed data centers. That makes EOS, MRC, SRv6 and the broader system-level architecture more valuable as networks become more complex. Scale-out is already material, scale-across is emerging as a potentially very large new routing-plus-switching market, and front-end networking should benefit as inference and agentic workloads place increasing demands on conventional data-center and enterprise networks.

The opportunity also becomes broader as accelerator architectures diversify. ANET has relatively limited scale-up exposure when NVIDIA supplies the full NVLink stack, but its addressable market expands with AMD MI accelerators, Google TPUs, custom hyperscaler silicon, inference accelerators and open Ethernet-based architectures. Bulls also see a customer diversification story: Microsoft and Meta remain major customers, but one or two additional 10% customers could emerge, while enterprise, campus, international and routing provide growth outside the largest cloud accounts.

This quarter materially strengthened that thesis. Revenue accelerated to 38% growth, Q3 implies 43%, FY26 was raised to 40%, Etherlink exceeded 100 customers and the core front-end and enterprise businesses contributed meaningfully to the raise. Supply is no longer an absolute ceiling, while ANET preserved roughly 50% OM despite component tightness. Product deferred revenue rising by another $600M alongside a substantial revenue beat suggests there is still demand waiting to convert, rather than the beat merely pulling revenue forward.

**The general bear case** is that ANET remains highly exposed to a small number of hyperscalers whose AI capex, architectures and purchasing decisions are outside its control. Microsoft and Meta together are still extremely important, and a capex slowdown, customer-specific deployment delay or share loss could have an outsized effect. Bears also argue that NVIDIA’s vertical stack, white-box hardware, proprietary hyperscaler networking and lower-cost open networking software could limit ANET’s share, particularly in scale-up where the company admits it has little participation today.

ANET is also committing substantial capital and balance-sheet capacity to secure components before the revenue is recognized. Purchase commitments have risen to $9.7B, inventory is elevated and deferred revenue depends on customer-specific acceptance clauses. If AI demand moderates or deployments move right, those commitments could become a working-capital and margin issue rather than a competitive advantage. The delayed timing of pricing benefits creates another risk if memory, silicon, optics or tariff costs rise faster than ANET can pass them through.

There were still several bear talking points in the quarter. The formal AI and campus targets did not increase despite the $1.1B overall raise, leaving room to question how much of the upside came from AI back-end share gains versus core shipments and deferred-revenue conversion. Billings growth slowed sharply to 16% from 54%, industry supply constraints are expected to persist through 2028, customer visibility is only two quarters and scale-up remains a selective 2027/2028 opportunity rather than a near-term revenue driver. At nearly 50x CY27 EPS street estimates, the stock still requires unusually clean execution although bulls will push back this still trades at 30x

### **ALAB -1.5%: Another big beat-and-raise, with 3Q revenue/EPS guidance 32%/47% above Street and above bogeys ($550M vs bogeys of $475M) as Scorpio X entered high-volume production one quarter early; the debate has shifted from whether the ramp is happening to how repeatable it is given AMZN concentration and valuation and concerns**

Overall solid print even against high expectations although seems like positioning even more crowded than we expected. Not much to pick at other than GM for Q3 which was guided down although still above expects. Long term questions still remain (see bull vs. bear debate).

### **The #s:**

**2Q:** Revenue $392.4M, +104.5% y/y (last q +93.4% y/y), +27.3% q/q vs Street $360.9M, +88.0% y/y; non-GAAP EPS $0.80 vs Street $0.69.

Margins: Non-GAAP GM 73.7% vs Street 72.8%; non-GAAP OM 39.1% vs Street 36.8%.

**3Q:** Revenue guidance midpoint $550M, +138.5% y/y and +40.2% q/q vs Street $417M, +~81% y/y; non-GAAP EPS midpoint $1.19 vs Street $0.81.

### **Key Takeaways:**

-

**Scorpio** Scorpio should become ALAB’s largest product family in 3Q, one quarter earlier than previously expected, with X-Series revenue overtaking P-Series. The lead 320-lane Scorpio X program is entering high-volume production, as the AWS Trainium 3 scale-up fabric; additional customers should begin contributing revenue around year-end, with more meaningful volume in 2027.

-

**Margins** The mix shift toward Scorpio is lowering GM, but the earnings model is showing substantial operating leverage. 2Q GM of 73.7% beat Street by roughly 90bps, while 3Q GM is guided to 72% vs Street around 71.7%; meanwhile, 3Q OM should expand to approximately 43% vs Street near 38%, despite the sequential GM decline. The callback suggested Scorpio margins should be modeled near the corporate average and company GM around 70% exiting 2026 and into 2027, as ALAB prioritizes socket capture over maximizing near-term percentage margins.

-

**Content** Mgmt framed the total ALAB opportunity as “multiple thousands” of dollars per XPU, with the Scorpio X switch itself moving above $1,000 as lane count, attach rates, optical and additional signal conditioning increase. The callback added an important nuance: current Scorpio X ASPs are not yet at roughly $1,000 but are trending toward it, so the headline should be read as an open-ended future-generation content trajectory rather than the current blended ASP.

-

**Aries** Aries posted record revenue and looks materially more durable than investors expected. PCIe Gen 5 is seeing a new inference-driven demand wave, while Gen 6 brings higher attach rates, higher ASPs and several new platforms; Gen 6 products across Aries and Scorpio already represented more than 50% of ALAB revenue, up from one-third in 1Q. Mgmt also expects the cycle to continue into PCIe Gen 7.

-

**COSMOS** The most strategically important incremental disclosure may have been that COSMOS is moving beyond fleet management, diagnostics and telemetry into the orchestration and data path. Mgmt said in-network compute and Hypercast can provide almost 2x improvement in collective operations, with XPUs directly interacting with COSMOS to run workloads. That materially strengthens the stickiness argument and makes the Scorpio socket harder to displace on future platforms.

-

**Portfolio** Mgmt talked up the growth story broadening beyond Scorpio and Aries. Taurus is moving from 400G into 800G, 1.6T and 3.2T, with the new 200G-per-lane portfolio doubling its TAM to more than $4B by 2030. Leo secured another US hyperscaler win and should ship standard and custom CXL controllers at two US hyperscalers in 2027, while fiber attach and NPO should begin contributing in 2027, followed by CPO in 2028+.

-

**Custom** Custom silicon is emerging as another 2027 growth vector, including NVLink Fusion-based hybrid rack designs and custom Leo controllers. Mgmt described the opportunity as multi-billion dollars over time, but initial shipments do not begin until 2027, so it remains future optionality rather than a near-term contributor.

-

**Concentration** The bear issue is not gone. The 3Q acceleration is still predominantly tied to the lead hyperscaler, and most other Scorpio X engagements remain in qualification or pre-production. The offset is that ALAB has more than 10 X-Series engagements and the lead customer itself spans multiple platforms rather than a single design, including training/inference, air/liquid cooling and different rack configurations.

-

**Competition** The core debate is whether ALAB can carry its PCIe incumbency into UALink and other scale-up fabrics against NVLink, Broadcom, Marvell and potential dual-sourcing. Mgmt’s answer is that its real-world Scorpio deployment experience, open-standard architecture and COSMOS integration create a meaningful lead, but customer diversification and merchant UALink adoption remain less proven than the current AWS program

### **Bull vs. Bear Debate:**

**The bull case** is that ALAB has crossed from being a high-quality PCIe retimer supplier into becoming a mission-critical AI fabric platform. Scorpio X is not merely a new product contributing incrementally; it changes ALAB’s content per accelerator, TAM and strategic role inside the rack. Scorpio becomes the largest product family one quarter earlier than expected, the lead 320-lane design is entering high-volume production, and the total content framework is moving from roughly hundreds of dollars to multiple thousands of dollars per XPU. Aries remaining at record levels at the same time materially improves the quality of the ramp because growth is not being created by an immediate collapse in the legacy franchise.

The second bull argument is that the initial AWS ramp can become a template rather than a one-off program. ALAB has more than 10 Scorpio X engagements, the lead customer has multiple platform variants, and COSMOS is now embedded in the orchestration and workload path rather than merely providing telemetry. That software integration raises switching costs and could help ALAB retain the socket as customers transition from PCIe-based fabrics toward native UALink. The 2027 pipeline also becomes more diversified through Leo CXL, Taurus 800G/1.6T/3.2T, fiber attach, NPO, custom silicon and additional Scorpio customers. The callback’s description of COSMOS as mission-critical and optics as four separate monetization paths reinforces the view that ALAB’s opportunity is broader than a single Trainium generation.

**The bear case** is that the current growth rate is being disproportionately created by a concentrated product transition at one customer. The 3Q guide is far more concentrated around the lead Scorpio X program than the broad-based 2Q beat, while most of the other 10-plus engagements remain in qualification or pre-production and may not reach true volume until 2027. The callback also clarified that current Scorpio X ASPs are not yet at the approximately $1,000 level, but are trending toward it, meaning investors should not apply the future content framework to every current unit.. Even within the lead customer, Trainium 4 and future generations could be contested by Marvell or other suppliers, and a hyperscaler may deliberately dual-source after ALAB captures substantial content. Multiple platforms reduce single-design risk, but they do not eliminate the economic dependence on one customer’s accelerator roadmap. Trainium 5 adds another risk because a material shift toward CPO could displace electrical Scorpio X unless ALAB’s optical roadmap is ready.

There is also substantial architecture and execution risk beyond the current program. UALink adoption is not guaranteed to occur on the timeline or physical layer ALAB expects; NVLink remains entrenched, Broadcom has formidable switch silicon capabilities, and Ethernet-based scale-up approaches may reduce the opportunity available to a native UALink switch. Aries could eventually face attach-rate normalization or share losses, while CXL, optics and custom silicon are still mostly 2027-plus opportunities. At the same time, company GM is expected to normalize toward roughly 70%, inventory and working capital have increased, and the stock leaves little room for a delayed qualification, weaker topology or lower lane utilization.

**What changed this quarter?**

The quarter largely answered the near-term question of whether Scorpio X would ramp: it entered volume production, Scorpio becomes ALAB’s largest product family one quarter early, and the 3Q revenue guide exceeded even most optimistic buyside bogeys.. Record Aries revenue, the 320-lane ramp and COSMOS moving into orchestration all strengthen the bull case. **However, the quarter did not answer the more important 2028 question: whether ALAB wins the Trainium 4 UALink switch and I/O-die opportunities, successfully participates in NVLink Fusion, and develops optical capabilities quickly enough for Trainium 5**

### **BKNG +9%: Clean 2Q beat, but the 3Q guide landed 2–4 points below Street (inline with bogeys) and FY bookings/EBITDA language softened as expected; stock up as expectations were low and core accommodation demand held up and Middle East-related pressure began easing in June and July. Stock leaned short going in so this better than feared**

### **The #s:**

**2Q **Revenue $7.352B, +8.1% y/y (last q +16.2% y/y) vs Street $7.191B, +5.8%.
Room nights 325M, +5.3% y/y (last q +5.9%) vs Street 320M, +3.8%; gross bookings $51.0B, +9.2% y/y vs Street $49.4B, +5.7%.
Adj. EBITDA $2.648B, +9.3% y/y, 36.0% margin vs Street $2.559B, +~5.7%, 35.6%; adj. EPS $2.54 vs Street $2.41.

**3Q** room nights guide of +3–5% and gross bookings/revenue/adj. EBITDA of +4–6% were below Street at +6.1%/+8.7%/+7.8%/+8.6%.

**FY26 G**r**oss bookings** high-single-digit growth, lowered from high single digits to low double digits; accommodation assumptions were broadly unchanged and the reduction was mainly flights.

**FY26 Adj. EBITDA** high-single-digit growth vs Street approximately $10.9B, +~9.7%; prior language called for EBITDA to grow “slightly faster” than revenue.

### **Key Takeaways**

-

**Demand** Travel demand remained resilient, with domestic and intra-regional travel materially stronger than long-haul international. U.S. room nights grew high single digits, Europe and Asia grew mid single digits, and domestic room nights globally grew high single digits, versus only slight growth in international room nights. Booking windows and length of stay shortened during the quarter but began normalizing in June, while July demand remained healthy.

-

**Middle East** The direct impact was slightly worse in 2Q than 1Q because weakness continued through April and May, but trends began normalizing in June and improved further in July. Middle East booker demand is now largely normalized, although inbound Middle East travel, elevated airfare, constrained airline capacity and weaker long-haul demand remain embedded through 3Q. FY guidance assumes seven months of Middle East impact, versus the prior assumption that disruption would fade around the end of 2Q.

-

**Cadence** The full-year outlook requires some acceleration in 4Q. Depending on where results fall within the ranges, the guide implies roughly 5–12% 4Q revenue growth, while one model estimates a roughly 10% revenue exit rate and nearly 12% EBITDA growth. Bulls will argue BKNG routinely guides conservatively and delivers larger H2 beats; bears will focus on the assumption that the Middle East impact largely clears by 4Q.

-

**Margins** 2Q adj. EBITDA margin expanded nearly 40 bps to 36.0%, 40 bps above Street, as fixed-cost leverage and customer-service efficiencies offset higher investment. The debate shifts to H2 because marketing expense grew 11% versus revenue growth of 8%, IT expense grew 20%, and the FY guide no longer explicitly promises EBITDA growth above revenue. The offset is an incremental $100M of transformation savings, lifting the annual run-rate target to $650M, although most of the new savings arrive in 2027

-

**AI** There is still no evidence of external AI disruption in the reported numbers. Paid and organic LLM referrals remain significantly below 1% of room nights and have not changed materially over the last few months or quarters. BKNG is participating in OpenAI’s CPC advertising test, is an early Google agentic-booking partner and believes its performance-marketing expertise can transfer to new channels. Internally, mgmt says AI already produces positive ROI, with customer-service cost per booking falling at a double-digit rate and AI spend still only a low-single-digit percentage of total technology spend.

-

**Direct** B2C direct mix remained stable in the mid-60% range and direct room nights continued to grow, despite modest SEO pressure from Google interface changes and AI Overviews. App mix and the share of room nights from higher-tier Genius customers were both in the high-50% range. Genius Levels 2 and 3 now represent more than 30% of active customers and generate a high-50% share of room nights, supporting the view that loyalty and direct behavior remain intact even if direct mix has stopped expanding materially.

-

**Connected Trip** Connected Trip transactions grew low double digits, more than twice overall transaction growth, and represented a low-double-digit percentage of Booking.com transactions. Merchant mix rose roughly four points to 73%, strengthening the payments and cross-vertical flywheel. However, the underlying verticals were mixed: flights grew only 3.7%, rental-car days declined 6.5%, and alternative-accommodation room nights grew 4%, below total room-night growth and likely below ABNB.

### **Bull vs. Bear Debate**

The central bull argument is that BKNG is still producing high-quality growth through an unusually noisy travel environment. Room nights, gross bookings, revenue, EBITDA and EPS all exceeded Street, despite a larger Middle East impact in 2Q than in 1Q. More importantly, the weakness was concentrated in long-haul travel and flights, while the economically more important accommodation business remained healthy. U.S. room nights grew high single digits, domestic travel remained strong across regions, ADRs rose 2% in constant currency and demand improved as the quarter progressed. With Middle East booker behavior largely normalized and 2027 set to lap the disruption, bulls see the current slowdown as cyclical rather than evidence that BKNG is losing relevance.

The strategic data also support the compounder thesis. Connected Trip transactions are growing more than twice as fast as total transactions, merchant penetration reached 73%, Genius Levels 2 and 3 are driving a high-50% share of room nights, and direct traffic remains stable despite SEO pressure. These initiatives can improve repeat behavior, payment economics and customer acquisition efficiency over time. The additional $100M of transformation savings, continued fixed-cost discipline and large repurchases give BKNG multiple ways to sustain mid-teens EPS growth even if room-night growth remains in the mid-to-high single digits.

Bulls also argue that the AI bear case remains theoretical rather than visible in the financials. LLM referrals are significantly below 1% of room nights and are not materially increasing, while direct room nights continue to grow. BKNG is not ignoring the transition: it is working with Google and OpenAI, building its own discovery products, and using AI internally to lower customer-service costs and improve software productivity. The bull outcome is that agentic platforms become another measurable performance-marketing channel, while BKNG retains the transaction through superior inventory, loyalty, payments, service and conversion expertise.

**The bear argument** starts with the fact that H2 expectations did go down. Every major 3Q growth guide was below Street, FY gross-bookings guidance was reduced, and the prior promise that EBITDA would grow slightly faster than revenue disappeared. The full-year math requires a meaningful 4Q reacceleration and appears to assume that most Middle East disruption clears after seven months. If elevated airfare, route constraints or weaker international demand persist, the 4Q and 2027 setup could prove less favorable than investors now expect.

There are also signs that growth is becoming more expensive. Marketing grew 11% against 8% revenue growth, IT expense grew 20%, and direct mix is stable rather than expanding. SEO remains small, but Google’s AI Overviews are already pressuring organic visibility. Alternative-accommodation room nights grew only 4%, below overall room nights and likely well below ABNB, while flights and rental cars materially missed Street. Bears can argue that Connected Trip looks better at the transaction level than within several of its underlying verticals, and that the 2Q margin beat may not represent the sustainable H2 trajectory.

The structural bear case remains AI distribution. Google and other AI platforms are moving from travel inspiration toward live rates and in-platform checkout, while hotel operators would prefer to replace high-cost OTA traffic with direct connections. Even if LLM traffic is currently negligible, greater agent adoption could raise customer-acquisition costs, lower take rates or shift leverage toward hotels and discovery platforms. BKNG’s B2B expansion adds another trade-off because B2B unit economics are lower than B2C and rapid B2B growth can cannibalize higher-margin direct business.

### **SHOP +22%: Huge quarter with strength across merchant sizes, channels and geographies, another quarter of >30% GMV growth, low 30s GMV guide better than expects, GP > 30% and meaningful operating leverage.**

Q3 revenue and gross-profit growth were guided above Street while FCF margins remain in the high-teens to low-20s despite rising cloud/AI infrastructure costs. AI engagement also accelerated, with traffic and orders from AI-originated searches up 3x and Sidekick conversations reaching 34M. Call ongoing

**The #s:**

**Q2 Revenue:** $3.58B, +34% y/y vs Street $3.45B, +29% y/y
**GMV:** $115.6B, +32% y/y (+30% FXN) vs Street $112.0B, +~27% y/y
**Gross Profit:** +31% y/y vs Street +~26% y/y
**MRR:** $221M, +19% y/y vs Street $218M
**Take Rate:** 2.41%, +1bp q/q vs Street 2.37%
**FCF:** $654M / 18.3% margin vs Street $547M / ~15.9%
**Op Margin:** 13.6% vs Street 12.5%
**Q3 Revenue Guide:** Low-30s% y/y, implying $3.70–3.78B vs Street $3.59B / +26.3%
**Q3 GP Guide:** Mid-to-high-20s% y/y, implying $1.74–1.78B vs Street $1.73B / ~25%
**Q3 FCF Guide:** High-teens to low-20s margin, implying $681–820M vs Street $647M / 17.4% margin

# **TECH RESEARCH/NEWS**

*Slow on the research and news front today…*

### **SPCX/CRWV/NBIS/ORCL: BNP Says SpaceX Could Tighten GPU Supply but Also Reinforces Premium Neocloud Economics**

BNP says SpaceX’s closer NVIDIA alignment and aggressive compute build could crowd neoclouds on GPU allocation and enterprise AI workloads as infrastructure deployments accelerate into 2027. BNP nevertheless says CRWV, NBIS and ORCL remain best positioned if supply tightens because of their strategic NVIDIA relationships, although BNP notes SpaceX could still receive priority allocations. BNP also says SpaceX’s commentary supports an AI compute shortage persisting through 2027 and potential Vera Rubin monetization of $30B–$50B per GW with sub-one-year payback, materially above BNP’s current $10B–$15B/GW assumptions and validating strong economics for providers that can consistently secure GPUs.

### **ChinaAI**: **China’s Open-Weight Models Will Be Spared US Safety Tests**

**[Bloomberg](https://www.bloomberg.com/news/articles/2026-08-05/china-s-open-weight-models-to-be-spared-us-tests-us-firms-told):**
**

The White House has told top US artificial intelligence companies that open-weight models being developed by their Chinese rivals won’t be subject to government testing under the Trump administration’s new AI safety framework, according to people familiar with the matter.

### **Memory: Citi Says Reported Rubin Ultra De-Spec Reflects Shortage, Not Demand Weakness**

Citi says recent concerns over lower HBM content per GPU are misplaced, as its supply-chain checks suggest persistent HBM shortages are forcing AI chipmakers to spread limited memory across more GPUs rather than cut aggregate demand. Citi notes a potential shift from 12-Hi/288GB Rubin to 8-Hi/192GB Rubin Ultra would reduce HBM per GPU, but says system-level HBM capacity could still rise 433% to 110.6TB as GPU count expands from 72 to 576. Citi also says lower HBM per GPU should increase KV-cache offload to server DDR5 and enterprise SSDs, creating additional demand across the broader memory stack.

### **Memory: HP, Asus and Acer begin using CXMT chips amid memory shortage**

[Nikkei](https://asia.nikkei.com/business/china-tech/hp-asus-and-acer-begin-using-cxmt-chips-amid-memory-shortage)
**

Leading PC makers HP, Asus and Acer have started to use small amounts of chips from China's ChangXin Memory Technologies amid an unprecedented memory shortage fueled by demand for artificial intelligence infrastructure, Nikkei Asia has learned.

The amount of CXMT chips used and the number of notebook models utilizing them are very limited as of now, as CXMT is prioritizing a large part of its production capacity for Chinese clients, such as Huawei, the people said. The notebook models that use CXMT chips are for sale in non-U.S. markets.

### **PANW: Barclays Raises PT to $370 as Positive Checks Support ARR Upside and a Path to 40% FCF Margins**

Barclays raises its price target to $370 and says strong firewall demand, healthy renewal trends and easier-to-close Mythos deals support upside to 4Q RPO and NGS ARR. Barclays expects FY27 NGS ARR growth of roughly 20% and believes management can guide at least in line with Street, while Barclays trims FY27 EBIT margin to 30% from 31% to reflect CyberArk and Chronosphere integration. Barclays says improving demand, MythoS-related security spend and acquisition synergies should still support a path toward 40% FCF margins in FY28.

### **ABNB: Hearing 3P saying Q3TD GBV tracking above consensus, driven by relative strength in both ADRs and Nights.**

### **AAPL: Apple is Scrambling for DRAM Weeks Ahead of Ultra, iPhone 18 Launch**

**[Culpium](https://www.culpium.com/p/exclusive-apple-is-scrambling-for?selection=16a229cc-06a8-4e64-8a4f-9149a15a4fab&r=5o29&utm_medium=ios&triedRedirect=true#:~:text=With%20less%20than%20six%20weeks%20until%20the%20expected%20debut%20of%20the%20foldable%20iPhone%20Ultra%20as%20well%20as%20the%20iPhone%2018%20and%20iPhone%2018%20Pro%2C%20assemblers%20are%20working%20with%20Apple%20to%20rush%20shipments%20of%20DRAM%20used%20in%20mobile%20devices):**
**

With less than six weeks until the expected debut of the foldable iPhone Ultra as well as the iPhone 18 and iPhone 18 Pro, assemblers are working with Apple to rush shipments of DRAM used in mobile devices and sturrling to get enough DRAM

### **Solar/GLW: US weighs polysilicon price floor, tariffs to counter China in solar and chips**

**[Reuters](https://www.reuters.com/world/china/us-weighs-polysilicon-price-floor-tariffs-counter-china-solar-chips-2026-08-04/):**
**

The Trump administration is preparing to set a price floor and impose tariffs on polysilicon and related products, according to four people familiar with the plan, putting a material critical to solar panels and semiconductors at the center ​of U.S. efforts to compete with China on artificial intelligence and energy.

The decision, expected to come later this month, aims to protect U.S. polysilicon factories owned by Hemlock Semiconductor and Wacker Chemie [(WCHG.DE), opens new tab](https://www.reuters.com/markets/companies/WCHG.DE) ‌from growing Chinese ambitions in the chip supply chain.

### **AI Tweets:**

*SSI model in August*


*Codex contnues to ramp:*



### **Leopold and Citadel**

Bloomberg: "Citadel’s flagship fund surged 5.9% in July, after the firm bought a discounted portfolio of artificial-intelligence stocks from troubled hedge fund Situational Awareness.

Before the deal, the Wellington fund at Ken Griffin’s firm had only been up 0.45% for the month through July 24. The transaction also helped boost Citadel’s year-to-date gains to 12%"

###
SPCX AMD ALAB ANET BKNG SHOP & More