On its 3Q26 earnings call, Amazon said it expects its cloud unit AWS to keep growing faster than 40% a year through 2028; this figure is a target/forecast, not an achieved result. The company also indicated that costs from its satellite internet project will hold back fourth-quarter profit. The source does not state the cause of the AWS growth target.
On Amazon's third-quarter 2026 earnings call, AWS reported revenue growth of 39.3% year-over-year in the third quarter of 2026, the cloud unit's fastest pace in nearly five years and its sixth straight quarter of acceleration. Segment sales reached US$46.0 billion, putting AWS at a US$184 billion annualized run rate, and operating income rose 55% to US$17.8 billion. The same call attributed AWS's choice to start selling its own AI chips to TSMC capacity. The source does not state the timing or implementation status of this decision.
AI chipsAmazon Web Services (AWS), TSMCDigiTimes Asia
Anthropic's "multi-platform resource pool" strategy, disclosed in its S-1 filing, is sending an important signal to the AI chip industry. The company uses Amazon Trainium, Google tensor processing units (TPUs), Nvidia GPUs, and AMD GPUs, integrating computing capacity from different partners into a unified resource pool through software and allocating it across chip platforms based on task requirements. The source states this strategy pressures Nvidia's pricing power.
AI chipsNvidia GPUsAMD GPUsAnthropic, Nvidia and 5 totalDigiTimes Asia
Amazon's third quarter showed its cloud expansion translating into stronger operating profitability, while the cash demands of building AI infrastructure kept rising. The result sharpens the company's central operating challenge: sustaining rapid AWS growth while bringing new capacity into service and maintaining the economics of its much larger retail business. The source does not state any specific component or price movement.
According to The Wall Street Journal, Apple, once one of the most powerful buyers in the NAND flash market, is losing some of its negotiating leverage as AI-driven demand gives memory suppliers greater pricing power. Meta, Microsoft, and Amazon have signed long-term supply agreements with Kioxia this year at higher prices, strengthening the chipmaker's negotiating position. Separately, South Korean outlet Economy Tribune reports that Apple has reportedly turned to a long-term NAND supply agreement, with market watchers identifying Kioxia as the likely partner; the terms remain undisclosed and this is a reported plan. The report adds that Apple may also pursue long-term agreements with Samsung Electronics, SK hynix, and Micron to secure future LPDDR5X and LPDDR6 supplies. According to TrendForce, memory's share of the 256GB iPhone Pro's bill of materials rose from around 10% a year earlier to 34% in 3Q26 and could exceed 40% in 1H27; TrendForce also estimates the iPhone 18 Pro's total BOM cost could rise 38% YoY, potentially forcing Apple to sacrifice margins to maintain shipments.
NAND flashLPDDR5XLPDDR6Apple, Kioxia and 5 totalTrendForcePrice Movement
Related manufacturers
Apple, Kioxia, Samsung Electronics, SK hynix, Micron
Microsoft and Nvidia publicly released the Surface Laptop Ultra on October 7, built around the application processor originally referred to as the N1X and now renamed the GB10. The chip was co-developed by NVIDIA (which contributed the Grace Blackwell GPU subsystem) and MediaTek (which supplied the multi-core CPU cluster and reportedly also handled full SoC integration duties); it was originally intended for—and may eventually still show up in—Arm-based Windows PCs, but repeated development hurdles reportedly delayed both SoC and system shipment plans, and lingering functional bugs precluded Windows compatibility (explaining the DGX Spark's Linux OS foundation). The machine comes in two versions sharing the same NPU core: 5120-core GPU with 18-core CPU, and 6144-core GPU with 20-core CPU; system memory ranges from 24 GB to 128 GB of LPDDR5x unified memory; SSD options start at 512 GB (PCIe Gen 4) with 1 TB and 2 TB (PCIe Gen 5) configurations; pricing is $2,599 baseline up to $5,899.99 top-end, with delivery promised 'later this year'; an RTX Spark-based developer platform at $5,999.99 and up is also available for pre-order. In the same period, Amazon launched its latest-generation Alexa-branded (no-longer-Fire) tablets in 8-inch, 11-inch and 12-inch 'Pro' flavors, running 'vanilla' Android with a return to the Google Play Store, priced from $299 up to $679.99 (not including cases or the high-end model's optional pen and keyboard accessories). Apple's week began Monday with a new-generation iPad mini (now with the A20 Pro, a slightly larger 8.5-inch display switched from LCD to OLED, priced from $699 and up; the 2024 entry-level 128 GB Wi-Fi-only model was originally $499 but was bumped to $599 in June of this year), Tuesday brought a 14-inch entry-level MacBook Pro now running the M6 SoC (versus last year's M5) plus a 24-inch iMac making a two-generation jump from the M4 to the M6, and today Apple unveiled its first touchscreen-augmented MacBook Pro laptops in both 14-inch and 16-inch displays with ProLED XDR panels and both M5 Pro and Max SoC options, priced from $2,899 to $14,549.
Nvidia RTX Spark N1X (GB10) application processorApple A20 Pro SoCApple M6 SoCApple, Microsoft and 6 totalEDN
Marvell is reportedly considering acquiring Lumentum to expand its product portfolio. Lumentum produces optical communication lasers and optical transceiver modules and currently has a market capitalization of approximately US$83 billion, compared with Marvell's US$221 billion; Lumentum has also attracted a US$2 billion investment from NVIDIA. A potential acquisition would further broaden Marvell's product portfolio, fill gaps in its optical component offerings, and strengthen its ability to compete with Broadcom. Bernstein analysts cautioned that the potential deal's size could raise concerns over financing and integration, noting Marvell may be unable to fund it entirely with cash; as of its latest quarter Marvell held approximately US$3.92 billion in cash and cash equivalents. Background: earlier this year Marvell acquired photonics company Celestial AI, gaining access to its Optical Multi-Chip Interconnect Bridge (OMIB) technology, and reached a chip agreement with Google granting Google the right to purchase up to US$12.2 billion worth of Marvell shares. The deal remains at a reported, considering stage and is not confirmed as agreed or completed.
NVIDIA's GeForce RTX 5080 has seen a significant increase in its listed price over the past few weeks, now costing nearly $2,000 at U.S. retailers. Cause: price hikes throughout the supply chain and the GPU's 16 GB of GDDR7 memory, which is currently in short supply, making the memory modules much more expensive. The card launched in early 2025 with an MSRP of $999; add-in card partners' custom solutions (new PCBs, enhanced cooling, and other features) gradually increased the price, and the issue escalated when memory supply became constrained. Across U.S.-based retailers like Best Buy, Micro Center, Amazon, Newegg, and B&H, the listing price now starts at nearly $2,000; on B&H, older listings show better pricing but are marked 'Temporary out of stock,' with only in-stock listings priced at $2,000 or more. In two months, listings have jumped from $1,200 to $2,000.
GeForce RTX 5080 GPUGDDR7 memoryNVIDIATechPowerUpPrice Movement
Anthropic disclosed details of its computing partnerships with Google, Amazon Web Services (AWS), Microsoft, Advanced Micro Devices (AMD) and other providers in its S-1 filing for an initial public offering (IPO), revealing distinct cooperation models and chip technologies across its partner ecosystem. Per the title, Anthropic's surging compute demand fuels ASIC prospects, favoring Broadcom over MediaTek.
On October 21, during GTC Berlin, Nvidia announced several multi-billion-dollar deals with six major supply chain partners — Ibiden, Unimicron, Samsung Electro-Mechanics, SPIL, Amkor and KYEC — across substrates, advanced packaging, semiconductor testing and optical engines, to expand production capacity and reserve capacity for its AI accelerators; expanded supply is to be available from 2027 through 2028, with commitments and production roadmaps stretching through 2028. For substrates, Nvidia plans to secure 2.5x dedicated substrate capacity over the next three years: it earlier signed a deal to provide a multibillion-dollar capacity reservation advance to support Ibiden's Ono Plant in Japan (targeting 3x the plant's current capacity); the Samsung Electro-Mechanics deal will support facility upgrades and new plants in Busan and Sejong, with Nvidia projected to account for a significant portion of its output by 2028; a capacity reservation agreement with Unimicron supports its substrate expansion in Taiwan, expected to start providing capacity by Q2 2027 and reach full capacity by mid-2028. For co-packaged optics, Nvidia signed with SPIL to expand optical engine production capacity 15x (1,400%) by mid-2027, with plans to double again by the end of 2028. With Amkor, Nvidia targets a more-than-300% increase in interposer wafer capacity by the end of 2028 through advanced packaging expansion in South Korea; Amkor confirmed the deal is worth $1.3 billion, bringing Nvidia's total investment across Amkor's Arizona and South Korea facilities to $3 billion. For testing, on October 13 Nvidia signed a deal to pay KYEC $1.1 billion to reserve testing capacity at KYEC's new $1.4 billion U.S. facility (high-volume production slated to begin in April 2027; Singapore high-volume production expected to begin February 2027), with testing capacity expected to double from calendar 2026 to 2027; KYEC currently handles nearly all of Nvidia's chip testing. Per its August 2026 disclosures, Nvidia had accumulated $279 billion in outstanding supply and manufacturing-capacity commitments by July 2026. All expansion figures are announced plans and targets, not yet realized.
substratesadvanced packaginginterposer wafersNvidia, Ibiden and 7 totalTom's HardwareCapacity Change
Anthropic's IPO filing reveals an infrastructure strategy extending far beyond conventional cloud procurement: the Claude developer has locked in hundreds of billions of dollars of compute capacity across Google, Amazon, Microsoft, Broadcom, and a growing roster of GPU infrastructure providers, spanning TPUs, GPUs, and dedicated data centers. This is a capacity commitment arrangement disclosed in the IPO filing; the source does not state the cause.
Nvidia has effectively sent JEDEC back to the drawing board with its HBM standard: Nvidia revolted against HBM memory controller functionality being in its GPU chip and, in its NVHBM custom HBM standard, moved it to the HBM base die unit, freeing up space for more GPU cores; the change is part of its August NVLInk Fusion announcement. Nvidia claims (a vendor claim, not a verified achieved result) that, compared with standard HBM4E, NVHBM delivers up to 30 percent greater memory bandwidth, 15 percent lower HBM power consumption, and frees up to 25 percent more area on its GPU compute die; this is specific to Nvidia and is not an industry standard. According to reported Morgan Stanley data, Nvidia will be responsible for purchasing 37 percent of the global HBM supply in 2027, so the three HBM suppliers — Samsung, SK Hynix and Micron — are supporting NVHBM and even let Nvidia dictate that they must use TSMC's 3nm foundry process; Reuters reported Samsung will accede to this request, but per Seoul Economic Daily, Samsung would still build JEDEC standard HBM in-house for other customers. Amazon's Annapurna Labs is also supporting NVHBM as an Nvidia NVlink partner. JEDEC has not issued any public comment about NVHBM; when JEDEC announced the HBM4 standard in April last year, the HBM subcommittee chair was Barry Wagner, Director of Technical Marketing at Nvidia. Beyond the stated reason of freeing GPU-core space and moving the memory controller, the source does not state another cause.
HBMHBM4EHBM4Nvidia, Samsung and 5 totalBlocks & Files
According to an RT NEWS item reposted by TrendForce, Amazon said that AWS large-scale compute capacity is fully booked for 2026, most of 2027 is sold out, and a considerable share of 2028 is committed. CEO Andy Jassy also said he does not expect Memory and power shortages to end in 2027. The source does not state the cause of the fully booked capacity or the persisting shortages; this is a statement of capacity/demand booking status, not a realized price change or confirmed production cut.
large-scale compute capacityMemorypowerAmazon, AWSTrendForce Official (Market Data)Capacity Change
Amazon plans approximately US$220 billion in capital expenditure in 2026, mainly for cloud and AI infrastructure. However, delays in data center power and construction are causing server racks to arrive before they can be deployed; the bottleneck is leaving expensive equipment in warehouses, preventing it from immediately generating computing capacity and revenue. The source does not state any further cause, and the plan remains planned rather than executed per the source.
cloud and AI infrastructureAmazonDigiTimes AsiaCapacity Change
Amazon plans to invest more than EUR160 billion (about US$17.9 billion) in Germany in 2027 to expand its cloud computing, artificial intelligence (AI), low-Earth-orbit (LEO) satellite network, logistics and media-content businesses, while establishing its second robotics and AI research laboratory in Berlin. This is a planned investment expansion, constituting a supply-side signal with indirect rather than direct price impact.
At PyTorch Conference 2026, held in San Jose, California, on Oct. 20 local time, Google and Amazon unveiled the open-source frameworks TorchTPU and TorchNeuron, respectively, designed to simplify migrating existing AI applications from Nvidia GPUs to their own chips and to challenge Nvidia's CUDA ecosystem dominance. PyTorch, developed in 2016 primarily by Meta's (formerly Facebook's) AI research lab FAIR and used by approximately 85% of AI application developers worldwide, was originally designed around Nvidia GPUs and the CUDA platform, making migration to other chips difficult. TorchTPU, developed by Google in collaboration with Meta, according to Google, allows more than 98% of existing PyTorch models to run on TPUs simply by changing the device string in the code to model.to("tpu"), with conversion taking from a few minutes to a few days, and Google said it enabled TPUs to operate as native devices on an equal footing with GPUs and CPUs at the core PyTorch level. AWS's TorchNeuron similarly allows PyTorch-based AI models to run directly on AWS's Trainium AI chips, requiring developers to change only a few lines of device settings, replacing Nvidia's cuda with AWS's neuron. Amazon has released TorchNeuron as a public beta through GitHub under the Apache 2.0 open-source license and is also contributing code directly to the main PyTorch open-source project. Previously, migrating AI applications optimized for Nvidia's ecosystem to Google's or Amazon's proprietary AI chips took at least several days and, in more complex cases, several months; the new tools effectively remove these time and cost barriers. As TorchNeuron, TorchTPU and similar tools gain adoption, dedicated AI chips are expected to compete primarily on underlying performance, power efficiency and pricing, though additional time will be needed to verify compatibility, performance and costs for individual models.
Nvidia CUDATorchTPUTorchNeuronGoogle, Amazon and 5 totalThe Elec
Related manufacturers
Google, Amazon, Meta, Nvidia, Amazon Web Services (AWS)
Amazon announced TorchNeuron, a native backend that allows existing PyTorch code to run directly on AWS Trainium chips with minimal code changes. The tool entered public beta today under the Apache 2.0 open-source license at the PyTorch Conference North America. The goal is to lower AI hardware adoption barriers and offer alternative compute choices.
Meta revealed this week that more than 11 percent of the GPUs at its Prometheus supercluster in New Albany, Ohio, are either dead or sitting idle, waiting to step in when a GPU fails. Prometheus is among Meta's largest compute deployments to date, powered by Nvidia's Grace Blackwell and (soon) Vera Rubin rack systems, and Meta says the supercluster will eventually scale to a gigawatt of capacity spanning about a million accelerators. The cluster is currently powered by a combination of GB200 and GB300 systems, each with 72 GPUs per NVLink domain, but only 64 are ever in use at any given moment, with the other eight essentially hot spares; Meta did this because a single failed GPU would take out an entire rack and interrupt the training run. Meta says that when it began the first phase of the Prometheus deployment it was experiencing roughly 1,000 interruptions per day for every 8,000 GPUs, and in the year since it has cut interruptions by 99 percent, to about 11 per day per 8,000 GPUs. Meta also says it achieved more than 85 percent goodput for pretraining and more than 90 percent for reinforcement learning on its latest model Muse Wonder, and as it works toward a million GPUs it hopes to reduce the interruption rate to just two per day for every 8,000 GPUs, or about 100,000 GPU-hours mean time between interrupts. The source does not state any implication for GPU market supply or demand.
According to a Wccftech interview with NVIDIA's Vera team reported on 2026-10-19, NVIDIA has chosen a monolithic compute die for its Vera CPU over the chiplet architectures used by major x86 competitors; Vera also features a custom LPDDR5X memory subsystem and NVIDIA's second-generation Scalable Coherent Fabric. As the report explains, unlike traditional cloud workloads that prioritize high core counts and throughput, agentic AI requires strong single-threaded performance, high memory bandwidth, and low latency to execute sequential tasks quickly and keep GPUs highly utilized. The monolithic design allows CPU cores to communicate without crossing boundaries between separate chiplets, reducing data movement latency. NVIDIA says Vera delivers 3.4 TB/s of bisection bandwidth, three times that of the leading x86 processor used in its comparison; Vera delivers 1.5 times the performance of its predecessor, Grace; and for agentic workloads involving sequential tasks such as Python scripting and code compilation, the company claims up to twice the performance of leading x86 processors, with a 1.6-times advantage on a geometric-mean basis. Vera also introduces NVIDIA's Spatial Multithreading technology. NVIDIA initially positioned its Grace CPU primarily as a host processor for GPU systems, but the report states that growing demand for agentic AI and reinforcement learning has prompted a shift toward standalone Vera CPU racks. Separately, according to the Central News Agency, NVIDIA Senior Vice President of Personal Computing Jeff Fisher denied during an interview in Taiwan rumors that NVIDIA plans to discontinue its flagship GeForce RTX 5090 graphics card, calling them untrue and stressing that the RTX 5090 remains a key product that NVIDIA will continue to supply; he acknowledged supply constraints are affecting the global industry, not just NVIDIA, but said the company is doing everything possible to bring products to market. TechPowerUp adds that the RTX 5090 remains difficult to find near its US$1,999 MSRP, with some listings on Amazon and Newegg exceeding US$8,000.
As reported by Tom's Hardware on 17 October 2026, after rumors that RTX 5090 production was stopping, Nvidia senior VP Jeff Fisher reportedly said in a purported post-event interview that the card has not been discontinued and will continue to ship. He reportedly stated: "The supply chain (capacity) is indeed somewhat constrained, but NVIDIA is doing everything it can to get the product on the market... production has not been discontinued, and NVIDIA will continue to ship," while noting other variants of the RTX 5090 (chip) exist. Earlier rumors held that, after Nvidia reportedly put the RTX 50 Super series on hold, production of the already-rare RTX 5090 was stopping, likely to use precious silicon manufacturing capacity for more AI accelerators. Retail spot-checks: RTX 5090s on Amazon sell for about $8,500 from third-party sellers; Newegg lists 5090s for about $7,300, likely from Chinese sellers judging by shipping time, with U.S. tariffs likely adding to the price; Best Buy has zero cards on offer; and as of this writing the card remains absent from Nvidia's own marketplace and Micro Center. Tom's Hardware received no reply from Nvidia for comment; the claims are reported second-hand, and production status is not officially confirmed.