❯ NVIDIA AI server systems up ~17%, adding $5 billion to the cost of a 1-gigawatt data center

[PRICE NOTICE] NVIDIA has notified major customers through server vendors that its flagship AI server system prices are rising roughly 17%, with the increase landing on machines shipping early next year and covering both the Vera Rubin and Grace Blackwell platforms. According to The Information, that single change adds at least $5 billion to the system procurement cost of a 1-gigawatt data center — before power, cooling, networking, construction, and financing are even counted in.

[COST DRIVER] This round of increases is not NVIDIA marking up prices — the company itself can no longer absorb the cost. Data from market research firm TrendForce shows server DRAM contract prices rose 53% to 58% quarter over quarter in Q2, with another 13% to 18% expected in Q3. Bloomberg earlier reported a more conservative figure, with multiple top-tier customers told the increase would be 15% or more; the two outlets’ numbers point in the same direction. Memory price inflation has now traveled from consumer-grade parts all the way up to rack-scale systems — the first time it has been written explicitly into NVIDIA’s price quotes.

[WHO PAYS] For the past two years, every hyperscale data center ROI projection has rested on the default assumption that unit compute prices keep falling year over year. A 17% increase overturns that curve in one move — and it lands after projects have already signed power agreements and locked in land. The first to be recalculated: cloud providers’ depreciation periods and rental pricing — the same GPU fleet has to amortize the extra cost, either by stretching lease terms or by raising per-hour rates. And downstream application companies that built their business models on inference costs continuing to fall may have to tear up next year’s gross margin models entirely.

▪ SIGNALMemory shortages have finally moved from supply-chain news to a deduction line item in every AI project’s financial model.

❯ Nvidia in Talks to Invest in Perplexity at Over $30B Valuation, After Weighing Tech Licensing and Talent Poaching

[DEAL TALKS] According to The Information, Nvidia is in talks to join a new equity round for AI search company Perplexity at a post-money valuation above $30 billion — up more than 50% from the roughly $20 billion round about a year ago — with the total round potentially reaching several billion dollars. The report also surfaces a detail worth a second look: Nvidia’s original plan was not an equity stake, but spending billions to license Perplexity’s technology and poach some core employees, before pivoting to direct shareholding.

[REVENUE CURVE] Underpinning that valuation is a steep revenue curve. Perplexity’s annualized revenue has climbed from under $250 million at the start of the year to more than $750 million — tripling within the year. Bezos is also behind the company. For Nvidia, Perplexity is just one move in a busy month: it has also backed three data-center power and land developers, while open-sourcing its own models for free on the side.

[DEMAND LOGIC] Strung together, these moves form a single playbook: use capital to lock in future chip demand. The abandonment of the licensing-plus-poaching plan shows Nvidia wants more than the technology itself — it wants Perplexity to keep operating as an independent, steadily purchasing marquee customer. The ones who need to redo the math are AI application-layer founders: when the most upstream chip supplier is also your shareholder, the valuation can be negotiated very high, but how much bargaining power and technological independence get diluted in tandem is a question that must be settled well in advance.

▪ SIGNALJensen Huang is using equity to lock in customers’ future orders ahead of time — far more cost-effective than selling a batch of chips.

❯ SoftBank Plans Record ~$6.3B Retail Bond, Japan’s Biggest Ever, to Fund OpenAI Pledge

[SIZE] SoftBank Group plans to issue ¥1 trillion (about $6.3 billion) in retail bonds — the largest retail bond ever from any Japanese issuer and SoftBank’s third this year. According to Bloomberg, the 7-year notes are expected to be priced September 4 with a guidance coupon range of 4.3% to 4.9%; proceeds will go toward AI-related investments and repaying older debt. The size is nearly double SoftBank’s record ¥600 billion offering from April 2025.

[DEPLOYMENT] SoftBank’s investment commitments to OpenAI have exceeded $60 billion, while it also accelerates its own data-center builds to expand compute. The question is where the money comes from. Three straight retail bond offerings are themselves a sign that banks have little appetite to fully absorb SoftBank’s AI exposure — the risk institutions won’t take has ended up on the counter of Japan’s individual investors. SoftBank shares drifted lower after the news. Its ¥600 billion April retail bond was a record at the time; four months later, that number has been nearly doubled by SoftBank itself.

[RETAIL] A 4.3% to 4.9% coupon is highly attractive in the Japanese market, especially for retail investors accustomed to near-zero interest rates. But buying this bond is essentially tying household savings to OpenAI’s commercialization progress, with SoftBank’s own leverage structure in between. Japan’s individual investors are, for the first time, standing this directly at the exposure end of the AI capex cycle — with far less information at hand than the banks that politely declined the deal.

▪ SIGNALWhen institutional money turns picky, retail savings accounts become the final link in AI infrastructure’s funding chain.

❯ Documents Show Meta Agent Hatch Launching as Early as Late August, New Model Watermelon Set for October

[TIMELINE] According to internal documents seen by The Information reporter Jyoti Mann, Meta’s consumer-grade agent Hatch — built to rival OpenClaw — is slated to launch in late August or early September, with the company’s newest model, Watermelon, scheduled for October. Meta’s stated rationale for building Hatch: OpenClaw may be hot in technical circles, but it is too complex for ordinary users. Hatch is designed to handle operations like shopping directly within Instagram, and Meta has tested it in simulated environments for DoorDash, Reddit, and Outlook.

[CONVERGENCE] Hatch has so far been trained on Anthropic’s Claude Opus 4.6 and Sonnet 4.6, with plans to switch to Meta’s in-house Muse Spark at official launch. Watermelon, though, is the heavier track — in July, Meta superintelligence lead Alexandr Wang told an internal all-hands meeting that Watermelon has matched OpenAI’s GPT-5.5 on internal benchmarks, with training compute an order of magnitude higher than the previous-generation Muse Spark. The specific benchmarks have not been disclosed to date.

[SEQUENCING] Put the two dates side by side, and Meta has set a schedule in which the model and the product backstop each other: Hatch seizes the consumer-agent entry point in September, then Watermelon swaps in the underlying capability in October. What takes the hit is Anthropic’s enterprise-customer base — a product trained on your model replaces you on launch day, and that script will likely play out more than once. Even as model vendors supply their biggest customers, they are simultaneously cultivating their own replacements.

▪ SIGNALTrain on someone else’s model, launch on your own — Meta switches between API customer and competitor with no transition at all.

❯ Musk’s First All-Hands at Cursor: Grok Is Behind, Anthropic Is Leading

[ALL-HANDS] After SpaceX completed its $60 billion acquisition of Cursor, Musk addressed the company’s entire workforce for the first time. Per The Information reporter Grace Kay, he said outright that Grok has fallen behind its rivals, that he is “not used to losing,” and that Anthropic is currently leading the race. He also offered a verdict: AI will eventually become uncontrollable by humans.

[NUMBERS] That wasn’t false modesty. On Artificial Analysis’s intelligence index, Grok 4.6 is essentially tied with OpenAI’s GPT-5.6 Sol Max, with both trailing Anthropic’s Fable 5 Max. At the meeting, Musk previewed Grok 4.7 for release in three to four weeks, claiming it would “surpass all existing models” after being fine-tuned on SpaceX’s massive data. With xAI already folded into SpaceX, Cursor’s addition means compute, proprietary data, and the coding entry point are now all under one roof.

[CONCESSION] A founder openly conceding defeat in front of a team he just bought for $60 billion is usually not an unguarded moment but a way of setting the tone: resource allocation will now revolve around “catching up,” not “holding ground.” For Cursor’s engineers, that one sentence settles it — the product roadmap will most likely yield to the pace of model catch-up. The ones who need to reassess are the developers who treated Cursor as a neutral tooling layer — it is now a catch-up component inside a model company. How much neutrality remains will be answered in three to four weeks, when Grok 4.7 ships.

▪ SIGNALThe first sentence bought for $60 billion was an admission of falling behind — Musk isn’t looking to soothe the team; he’s reordering priorities.

❯ Nvidia Groq 3 LPX Reaches Mass Production, Hits 3,400 Tokens/Sec in 100K-Token Long-Context Testing

[PRODUCTION & BENCHMARKS] According to Nvidia’s announcement, the dedicated inference accelerator Groq 3 LPX has entered full mass production. In Artificial Analysis’ standard benchmark, running open-source model Gemma 4 31B with 100K tokens of active context, it hit an output speed of 3,400 tokens per second — 4x the next-best platform under the same long-context conditions, and the fastest result ever recorded for this model. A single rack can hold up to 256 LPUs, paired with 128GB of high-bandwidth SRAM.

[EIGHT-MONTH SETUP] The chip is positioned as a complement to the Vera Rubin NVL72 platform, aimed at agent-style workloads that need to continuously produce tokens over long stretches. The more notable point is the timeline: per The Register, it has been eight months since Nvidia secured a non-exclusive technology license from Groq, and this is the first time Groq technology has appeared in an Nvidia rack-scale product. Cloud provider Nebius is the first deployer, offering it externally through its Token Factory service.

[INFERENCE SPLIT] Training and inference are being split into two independent hardware product lines, and per-second token output under long context has become an independently priceable metric. The first thing to loosen is the response-latency budget for agent products: an agent that must make dozens of consecutive tool calls was previously limited by generation speed to asynchronous interaction; with speed now at this magnitude, synchronous real-time product forms are tenable for the first time, and per-task inference costs are being recalculated accordingly.

▪ SIGNALPer The Register, that roughly $20 billion licensing deal has, for the first time, taken a shape that benchmark tests can record.

❯ Xiaomi Unveils Three Xuanjie Chips at Once; D100 Supports Local Deployment of 200B-Parameter Models

[CHIP TRIO] According to Xiaomi’s official disclosure at a technical briefing, the company unveiled three self-developed chips in a single release. The flagship SoC, the Xuanjie O3, uses a 3nm process, measures 133 square millimeters, and packs 24 billion transistors (up 26% from the previous generation). Its CPU is a ten-core architecture — six super-large cores plus four large cores — with a top frequency of 4.35GHz and a 60% performance improvement; its GPU is a 16-core design with an 85% performance boost. The O3 has entered volume production and will debut on the Xiaomi 18 Fold foldable next month.

[AUTOMOTIVE & EDGE] The other two chips are even more telling about the direction. The Xuanjie D100 is Xiaomi’s officially announced China-first 3nm high-compute AI chip for intelligent driving, pairing a 20-core CPU with a 16-core NPU and supporting up to 160GB of unified memory, enabling on-vehicle local deployment of large models on the scale of 200 billion parameters. Development is complete, with commercial deployment slated for next year. The Xuanjie O100, meanwhile, is a high-bandwidth accelerator chip purpose-built for on-device large models, using the industry’s first 6nm wafer-level vertical stacking and hybrid bonding process, with a bond pitch of 1.4 microns and 1.22TB/s of bandwidth — 16 times that of a traditional flagship phone — delivering on-device inference speeds of up to 330 TPS.

[MEMORY WALL] All three chips point to the same conclusion: the bottleneck for on-device large models is memory bandwidth, not compute. The O100 uses advanced packaging to push bandwidth up an order of magnitude, while the D100 uses 160GB of unified memory to lift the entire “how large a model can run inside a car” ceiling up a notch. If 200 billion parameters can remain local, then three things — cloud inference call volumes, data-compliance pathways, and the billing basis for subscription models — all have to be reframed accordingly. And that line was drawn by Xiaomi itself, not by a chip supplier.

▪ SIGNALXiaomi put its most aggressive chip in the car, not the phone.

❯ XPeng Robotics Business Raises Over $900 Million at $6.3 Billion Valuation; IDG Leads, Tencent and Alibaba Follow

[FIRST-ROUND RECORD] According to XPeng’s announcement, its robotics business has completed its first financing round, raising over $900 million at a post-money valuation above $6.3 billion — a new record for the largest single private financing round in China’s embodied AI sector. The round was led by IDG Capital, with Tencent and Alibaba joining as strategic investors; Gaorong Ventures, an early backer of Momenta and Zhiyuan Robotics, also participated.

[USE OF FUNDS] The use of proceeds is spelled out in unusual detail: robotics software and hardware R&D, training and fine-tuning of physical AI models, high-quality data collection, construction of end-to-end mass-production lines, and overseas expansion. XPeng’s previously stated target is to reach monthly production of 1,000 IRON humanoid robots by year-end, deploying them first in its own stores and industrial parks, with commercial sales and deliveries set for 2027, launching simultaneously in China and overseas markets.

[MASS PRODUCTION] A robotics company that hasn’t sold a single unit has secured a $6.3 billion valuation — the pricing basis is not revenue but the visibility of mass-production capability. The two numbers, 1,000 units per month and 2027 delivery, happen to be among the few commitments in this industry that outsiders can actually verify. What now has to re-queue is the fundraising cadence of China’s other humanoid robotics companies — now that the front-runner has lifted the single-round record to the $900 million scale, when mid-tier players return to valuation talks, the first question they face will be: where is the production line, and what is the monthly output?

▪ SIGNALThe valuation anchor for embodied AI is shifting from research papers and demo videos to production lines and delivery timelines.

❯ ByteDance Folds Trae and Coze Into Doubao, Launches Doubao Work This Week to Take On Tencent WorkBuddy

[ORG FIRST] According to Bloomberg, ByteDance is folding the entire teams behind AI coding platform Trae and agent-building platform Coze into Doubao, and plans to launch Doubao Work as a standalone app as early as this week—a direct challenge to Tencent’s WorkBuddy, currently the most-used office AI tool in China.

[CONSOLIDATION] This is a textbook “bet on a single super-app” move. Trae launched domestically in March 2025 as one of the first homegrown AI coding tools to integrate Doubao 1.5 Pro and DeepSeek models; Coze is ByteDance’s agent-building platform launched in 2024. The developer audiences they target barely overlap with Doubao’s consumer base. Alibaba is doing the same, folding its AI tools into the office-focused Qwen Work.

[THREE-WAY RACE] China’s consumer-grade AI assistants have become so cutthroat that products are indistinguishable; office scenarios, by contrast, still offer clear willingness to pay and data moats. The first to feel the pressure is Tencent WorkBuddy’s usage lead—ByteDance has always used consumer-grade traffic density to go after rivals’ enterprise share. For enterprise buyers, the real cost is lock-in depth: once coding, agents, and office assistants are folded into one app, switching is no longer as simple as changing software.

▪ SIGNALByteDance moved the organizational structure rather than product features—a sign it sees office AI as a battle requiring concentrated firepower.

❯ Alibaba Releases Video Generation Model Wan3.0, Turning Documents, Spreadsheets, and Slides into 30-Second Videos

[LAUNCH & CAPABILITIES] Alibaba has officially released its video generation model Wan3.0, capable of generating up to 30 seconds of video per run at up to 1080p resolution — twice the duration of the previous generation. Its most distinctive feature is on the input side: it directly accepts DOC, XLS, PPT, PDF, and Markdown files, and can also read web pages, turning documents and tables into finished clips. The model opened public beta on August 6.

[PRICING & TIMING] According to the official pricing on Alibaba Cloud’s Bailian platform, API rates are $0.05 per second for 480p, $0.10 per second for 720p, and $0.20 per second for 1080p. Since the public beta began on August 6, it has already been used in short-drama and film/TV production, advertising and marketing, cultural-tourism promotion, and music videos. The release timing is also worth noting: it came the day after Alibaba completed an approximately $10.2 billion share placement in Hong Kong, with the proceeds explicitly earmarked for AI infrastructure.

[ENTRY SHIFT] The move from prompts to document input changes where these tools plug in — for the first time, video generation can be attached directly to a company’s existing document workflow, with no need to write a description first. The real bottom line lands on the outsourcing budgets of marketing and content teams: whether a quarterly report or a product manual can be turned directly into a finished video determines whether that money is saved or simply becomes one more subscription.

▪ SIGNALOnly when the input format shifts from prompts to PPT does video generation truly enter a company’s workflow.

❯ XPeng H1 2026 Net Loss Widens 173% to 3.121 Billion Yuan; R&D Up 39% Toward Physical AI

[WIDENING LOSS] XPeng Group reported total operating revenue of 32.777 billion yuan for H1 2026, down 3.8% YoY; net loss attributable to shareholders was 3.121 billion yuan, widening 173.35% YoY. But gross profit in the same period rose 20.25% YoY to 6.766 billion yuan, lifting overall gross margin to 20.6%, up 4.1 percentage points YoY — revenue is falling while gross margin is rising, and that divergence is the entry point to reading this report.

[SALES & R&D] The report shows H1 deliveries of 165,977 vehicles, down 15.8% YoY, and vehicle sales revenue of 28.05 billion yuan, down 10.3% YoY. The real support for gross margin came from another line: services and other business revenue of 4.73 billion yuan, up 67.1% YoY. Meanwhile, R&D spending hit 5.82 billion yuan, up 39% YoY, explicitly earmarked for new models as well as Physical AI and humanoid robots. At period end, cash on hand was 40.48 billion yuan, headcount at 20,632, of whom more than 8,700 were in R&D.

[LOSS NATURE] Pair this report with the same-day $900 million funding round for the robotics business, and the math becomes clear: fewer cars are being sold, but profit per car is up, and both the savings and the capital raised are flowing into robots. Q3 guidance calls for deliveries of 115,000 to 121,000 vehicles and revenue of 21.7 billion to 23.4 billion yuan. Valuing this company on deliveries will increasingly fail — the bulk of R&D spending is going into Physical AI, a cost that won’t appear on any near-term delivery statement, yet will materially shape next year’s pace of cash burn.

▪ SIGNALA carmaker that attributes the source of its losses to R&D investment — readers must first determine whether this is bleeding out or swapping positions.

❯ Smart Ring Maker Oura Plans US IPO to Raise Up to $3 Billion at Over $16 Billion Valuation

[TIMING & SCALE] According to Bloomberg, smart ring maker Oura and certain existing shareholders are seeking to raise up to $3 billion through a US IPO at a valuation of over $16 billion, with the deal potentially completing as soon as September. That is nearly 50% above the $10.9 billion valuation implied by last September’s $875 million Series E. The company confidentially filed its listing application in May, with underwriters including Goldman Sachs, Morgan Stanley, JPMorgan, Allen & Co., and Jefferies.

[EXIT] The report specifically notes that existing investors are expected to sell a significant portion of their shares in the offering; terms are still under discussion and could change. That detail makes the nature of the deal clear: less a hardware company raising money from the market to expand production than a secondary exit that has been queued up for a long time. Oura closed its $875 million Series E last September, led by existing investors; the company then confidentially filed this May — eight months later.

[REPRICING] A $16 billion valuation for a ring maker rests not on hardware gross margins but on subscription renewal rates and the long-term value of health data. Once public, those two metrics will be disclosed quarterly. The first to be repriced is the entire wearable health sector — Oura is the first of this cohort to truly reach the public market, and its P/E ratio will directly set the valuation ceiling for every comparable company that follows.

▪ SIGNALExisting shareholders choosing to sell at this price is itself a statement on the $16 billion figure.

❯ Pinduoduo Q2 Revenue Misses at RMB 112.36B; Net Profit Down 12% YoY but Beats Expectations

[MISS & BEAT] Pinduoduo’s Q2 total revenue came in at RMB 112.36 billion (about $16.72 billion), up 8.1% year over year, below the analyst estimate of RMB 116.35 billion; net profit was RMB 27.18 billion (about $4.04 billion), down 12% year over year, but above the market consensus of RMB 24.4 billion. According to Wall Street Journal reporter Tracy Qu, the company attributed the pressure to intense competition in the Chinese market and changes in the overseas regulatory environment.

[TWO-FRONT SQUEEZE] The earnings report and analyst notes show that on the domestic front, weak consumer confidence, employment expectations, and the drag from real estate kept shopping intentions conservative — even 618 failed to reverse the trend, and the e-commerce price war directly thinned profit margins. Overseas, Temu has taken the double hit of US tariffs on Chinese goods and the removal of the duty-free exemption for low-value parcels, pushing up shipping and compliance costs and forcing some merchants to raise prices. Advertising revenue grew just 3.8% in the same period.

[DECELERATING ENGINE] Pinduoduo has been one of the few Chinese internet companies able to sustain high growth in recent years; the 8.1% figure says that era is over. The first to rebuild assumptions should be e-commerce platforms on the same track: when even the industry’s fastest grower is down to single digits, the price war is no longer a way to grab share, but a cost item everyone has to bear.

▪ SIGNALOnce growth falls to single digits, the price war shifts from an offensive strategy to a war of attrition no one can afford to exit.

❯ Geely’s 500Wh/kg Solid-State Battery Slated for 2027 Multi-Brand Pilot, Volvo Viewed as Candidate

[TIMELINE] Geely says it is accelerating solid-state battery commercialization, planning pilot deployment across multiple brands in 2027, with Volvo viewed as one of the leading candidates for first adoption. The cell’s energy density reaches 500Wh/kg — nearly double that of existing lithium iron phosphate batteries — giving equipped models an expected range of over 1,000 km and a service life exceeding 1 million km. It also outperforms liquid batteries in safety, lightweighting, and charge/discharge efficiency.

[CONSTRAINTS] Mass production still faces hard constraints. The 2027 pilot is only small-batch testing — consumers won’t see production models in the near term, and the rollout timeline hinges on supply-chain maturity. Geely Holding’s brand portfolio spans seven or more tiers, covering Geely, Zeekr, Lynk & Co, Volvo, Polestar, Lotus, and smart. The company has yet to announce the first test brands and initial markets, and the industry broadly expects a phased rollout. Geely previously disclosed that the cell entered real-vehicle validation in 2026, with mass production timing depending on the supply chain.

[WEIGHT] What 500Wh/kg truly rewrites is not the range number but the weight and volume budget of the battery pack. At the same range, the pack weighs half as much, fully freeing the design space for the chassis, suspension, and vehicle structure — and the vehicle cost structure shifts along with it. The first thing to feel the strain is the vehicle platform development schedule: a platform takes three to four years from definition to mass production, and given that 2027 is only small-batch testing, the next-generation platform now has to decide whether to be engineered around liquid or solid-state physical parameters.

▪ SIGNALAn entire supply chain sits between pilot and mass production, and that time gap decides who is telling stories and who is scheduling production.

❯ Keep H1 Revenue RMB 825 Million, Adjusted Net Profit RMB 5.88 Million, Proprietary Sports LLM Moves into Core Scenarios

[SLIM PROFIT] Keep released its 2026 interim results, with H1 revenue of RMB 825 million and adjusted net profit of RMB 5.88 million. Two user-side metrics are improving: ARPU grew 21.3% year-on-year, and monthly active users’ average monthly exercise duration rose 15.3% — more people are paying, and they are staying longer.

[OWN BRAND] The financials show that hardware is carrying the revenue mix. Own-brand fitness product revenue rose 21.7% year-on-year to RMB 483 million, with gross margin up to 40.1%, and overseas revenue exceeding RMB 22 million. On the AI front, the self-developed sports-health LLM Keepace.ai continues to roll out into core App scenarios, while also exploring enterprise-facing capability exports.

[5.88M's WEIGHT] An adjusted net profit of RMB 5.88 million is almost nothing for a listed company, but it is the one figure in this report that proves this model can sustain itself at the current scale. The key is the causal link between ARPU and AI implementation: if the 21.3% growth in revenue per user genuinely comes from higher willingness to pay driven by AI courses and personalized training, then Keep’s story can continue; if it is just a price adjustment, next year’s interim report will reveal that.

▪ SIGNALA company has just proven it can avoid losing money; next it needs to prove this profit is replicable.