AI chips coverage across GPUs, accelerators, custom silicon, memory shortages, foundries, export controls, and the semiconductor supply chain behind AI compute.
Topic brief
Brief updated Aug 4, 2026
AI chips are the specialized processors and memory that make modern machine learning economically possible. The category spans training and inference accelerators from NVIDIA, AMD and Google, custom silicon programs at the largest cloud buyers, a field of inference-focused startups pursuing wafer-scale, chiplet and edge designs, the high-bandwidth memory and advanced DRAM that feed those accelerators, and the foundries, lithography suppliers and packaging houses that manufacture them. Because compute is the binding constraint on AI progress, the silicon layer sits underneath every model and product in the industry.
For practitioners, chips determine what is possible and what it costs. ML and infrastructure teams care about accelerator availability, memory bandwidth, interconnect topology and the price-performance of inference, which increasingly dominates real workloads. A related distinction now shows up directly in hardware choices: prefill, the compute-heavy stage that builds the KV cache, and decode, the memory-bandwidth-bound stage that emits tokens, have different ideal silicon, which is why heterogeneous racks are becoming a serious design pattern. Procurement teams track rack costs, lead times and the memory market, because an HBM shortage or a delayed rack system can reshape a plan. Software portability matters just as much as raw silicon, since a second source is only useful if the serving stack, kernels and numeric formats actually follow the workload across vendors.
The market has three defining features. NVIDIA remains dominant but is no longer unchallenged, facing rack-scale competition from AMD, custom accelerators at its largest customers, and inference specialists targeting the decode phase. Memory and packaging are the swing factors, with AI demand pulling DRAM away from consumer devices and advanced packaging capacity gating how many accelerators can actually be built. And chips are deeply geopolitical, with export controls, national industrial policy and sovereign compute programs shaping who can buy and build what.
The most concrete development is that the alternative to a single-vendor accelerator plan now has dates attached to it. AMD launched Helios at its Advancing AI event on July 23, a rack-scale system combining 72 Instinct MI455X GPUs with 18 sixth-generation EPYC Venice CPUs, Pensando networking and ROCm; AMD says it is in production, Reuters reports shipments near the end of the third quarter, and OpenAI expects to bring systems online beginning in the fourth quarter. Anthropic said on July 22 it will deploy up to 2 gigawatts of Instinct MI450-series GPUs in Helios racks with the first gigawatt planned for the first half of 2027, against an AMD commitment to a future equity investment of up to $5 billion that Reuters reports is tied to deployment milestones. Microsoft and AMD announced on July 20 that Azure will add three AMD-powered VM families, with Helios shipments to Microsoft beginning in the second half of 2026, and AMD and Cerebras announced a joint platform on July 23 that runs prompts and large context windows on Helios while Wafer-Scale Engine systems handle latency-sensitive token generation, targeted for Cerebras Cloud in the second half of 2026. Custom silicon is moving in parallel: Fubon Research projects Google could deploy 12 million to 15 million ninth-generation TPUs in 2028, near its 12.4 million forecast for Nvidia data center GPUs, though Google has confirmed neither the target nor TPU v9 specifications, and The Information reports an experimental Google chip called Frozen v2 that would embed parts of Gemini's architecture in silicon with internally projected efficiency of six to ten times more tokens per unit of power than Google's newest TPUs. The practical reading for a buyer is that second-sourcing is now a scheduling exercise rather than a hypothetical, but AMD's performance comparisons are vendor estimates and none of these programs has published reproducible workload benchmarks or software-portability evidence.
The binding constraint, meanwhile, keeps moving further upstream into wafers, lithography, packaging and memory. TSMC committed an additional $100 billion to Arizona manufacturing and packaging, taking its reported U.S. investment pipeline to $265 billion, and CFO Wendell Huang described demand as a multi-year mega trend while the company converts 5-nanometer capacity toward 3-nanometer production. ASML reported second-quarter net sales of EUR 9.3 billion and raised full-year guidance to EUR 43 billion to EUR 45 billion, attributing the improved visibility to AI-related investment in advanced logic and memory. Micron raised planned U.S. investment above $250 billion through 2035 with a separate commitment of up to $3 billion for supply chains, SK Hynix raised $26.5 billion in its July 10 Nasdaq debut to fund fabrication and advanced packaging, and Doosan agreed on July 31 to buy SK Inc.'s 70.6% stake in silicon-wafer maker SK Siltron for 2.3 trillion won. That capacity race has a visible cost elsewhere: Counterpoint Research estimates global smartphone shipments fell 11% year over year in the second quarter of 2026, the weakest second quarter since 2013, as memory suppliers favored higher-value AI data-center demand, and expects a roughly 14% full-year decline with the shortage possibly persisting into 2027. Around the edges sit risks that do not appear on a chip roadmap: China imposed a temporary helium export ban on July 10 affecting a gas used in wafer cooling, plasma etching and lithography support; the Federal Reserve named AI-related demand for semiconductors and data-center infrastructure as a near-term price pressure with May total PCE inflation at 4.1% and core at 3.4%; and export policy keeps shifting, with the United States making approved Emirati buyers eligible for advanced computing items without individual licenses while Beijing weighs limited H200 approvals for Alibaba, ByteDance and DeepSeek.
The nearest checkable milestones belong to AMD: whether Helios shipments land near the end of the third quarter, whether OpenAI brings systems online beginning in the fourth quarter, whether Microsoft's Helios shipments start in the second half of 2026 and turn into dated Azure availability and pricing, whether the AMD-Cerebras platform reaches Cerebras Cloud in the second half of 2026 with end-to-end latency, concurrency and energy measurements behind the cited up-to-5x tokens-per-watt figure, and whether Anthropic's first gigawatt of Instinct MI450-series capacity arrives in the first half of 2027 and triggers AMD's up-to-$5 billion equity commitment. On custom silicon, the open questions are whether Google ever confirms TPU v9 volumes or specifications against Fubon's 12 million to 15 million projection, whether the reported Frozen v2 project becomes an announced product or stays a trial, whether Tesla identifies the AI hardware company it acquired for $1.95 billion in the second quarter with $1.73 billion tied to service or deployment milestones, and whether Microchip's Hailo acquisition closes near the end of the quarter ending September 30. Upstream, watch whether Doosan's 2.3 trillion won SK Siltron purchase clears regulatory review and financing, whether TSMC's Arizona ramp and 5-nanometer-to-3-nanometer conversion hold, whether ASML's EUR 43 billion to EUR 45 billion guidance survives order timing and export controls, and whether Counterpoint's expected roughly 14% full-year smartphone decline and its warning that the memory shortage could persist into 2027 play out. On policy, the unresolved items are how long China's temporary helium export ban lasts, what H200 volume Beijing finally allows Alibaba, ByteDance and DeepSeek, whether license-free UAE access holds under its anti-diversion and matching-investment commitments, and whether South Korea's green-paper process on sharing AI chip gains produces any binding instrument.
| scope | open items | stated timing | customer or partner |
|---|---|---|---|
| Helios rack-scale systems combining 72 Instinct MI455X GPUs with 18 sixth-generation EPYC CPUs, Pensando networking and ROCm | AMD's performance comparisons are vendor estimates, so reproducible workload benchmarks, software compatibility and deployment evidence are still outstanding | AMD says Helios is in production; Reuters reports shipments near the end of the third quarter, and OpenAI expects to bring systems online beginning in the fourth quarter | OpenAI |
| Up to 2 gigawatts of Instinct MI450-series GPUs in Helios systems, plus a multi-year engineering collaboration around Claude and ROCm | AMD committed to a future equity investment of up to $5 billion that Reuters reports is tied to deployment milestones | First gigawatt planned for the first half of 2027 | Anthropic |
| Three VM families: ND MI455X v7 for large-scale AI inference, HDv2 for data pipelines and agent coordination, and HXv2 for chip design and technical computing, spanning accelerators, EPYC Venice CPUs, Pensando networking and Azure Boost | Deployment size, Azure availability dates and pricing were not disclosed | AMD says Helios shipments to Microsoft begin in the second half of 2026 | Microsoft Azure |
| Joint inference platform pairing Helios racks for prompt and large-context processing with Wafer-Scale Engine systems for latency-sensitive token generation | The companies cite up to 5x higher tokens per second per watt, but published reporting lacks additional performance data and interconnection details | Targeted availability through Cerebras Cloud in the second half of 2026 | Cerebras |
There is one with dates, though not yet with independent benchmarks. AMD launched Helios on July 23, a rack combining 72 Instinct MI455X GPUs, 18 sixth-generation EPYC Venice CPUs, Pensando networking and ROCm, and says the system is in production. Reuters reports shipments near the end of the third quarter. Named offtake includes OpenAI from the fourth quarter, Anthropic with up to 2 gigawatts of MI450-series capacity from the first half of 2027, and Microsoft Azure with Helios shipments starting in the second half of 2026. AMD's performance comparisons are vendor estimates, so reproducible workload benchmarks and software compatibility remain the critical next tests.
A substantial share. AMD reported a 46% share of data center CPU revenue on July 23, according to Wccftech, and estimates its total compute addressable market across server CPUs, AI accelerators and rack-scale systems could reach $2 trillion by 2030; earlier coverage cited a separate AMD forecast that the server CPU market alone could exceed $120 billion by 2030. AMD also extended the EPYC roadmap through 2030, confirming Zen 7-based Florence, Ferrara and Fidenza for 2028 and Zen 8-based Ravenna for 2030, with Florence reported to add ACE AI Compute Extensions and next-generation MRDIMM and LPDDR support. Core counts and performance targets are undisclosed.
Fubon Research projects Google could deploy 12 million to 15 million ninth-generation TPUs in 2028, close to its 12.4 million forecast for Nvidia data center GPUs, though Google has not confirmed the target or disclosed TPU v9 specifications and unit comparisons do not establish equivalent delivered compute. The Information separately reports Google is developing Frozen v2, an experimental server chip embedding parts of Gemini's architecture in silicon, with internal projections of six to ten times more tokens per unit of power than Google's newest TPUs; it is not an announced product. Broadcom's custom-ASIC business is tied to large cloud customers and it has announced a strategic collaboration with OpenAI for accelerator and networking systems. Tesla disclosed in a July 23 filing that it acquired an unnamed AI hardware company in the second quarter for $1.95 billion in stock and equity awards.
Consumer devices are absorbing it first. Counterpoint Research estimates global smartphone shipments fell 11% year over year in the second quarter of 2026, the weakest second quarter since 2013, as memory suppliers favored higher-value AI data-center demand and handset component costs rose. Counterpoint expects a roughly 14% full-year shipment decline and says the shortage could persist into 2027. On the supply side, SK hynix reported preliminary second-quarter operating profit of 60.5 trillion won at a 76% margin, said HBM4 mass shipments began in the quarter, and raised $26.5 billion in a July 10 Nasdaq debut to fund fabrication and advanced packaging.
TSMC committed an additional $100 billion to Arizona manufacturing and packaging, bringing its reported U.S. investment pipeline to $265 billion across a buildout intended to total 12 facilities, and is converting 5-nanometer capacity toward 3-nanometer production. Micron raised planned U.S. investment above $250 billion through 2035, with a separate commitment of up to $3 billion for supply chains including support for GlobalWafers in Texas, and a long-term goal of producing 40% of its DRAM in the United States. ASML raised full-year sales guidance to EUR 43 billion to EUR 45 billion with a 54% to 56% gross-margin range, citing AI-related investment. Flex and Cerebras added Milpitas lines expected to support roughly a sevenfold increase in CS-3 production through 2026.
The United States eased controls for the United Arab Emirates, making the UAE government and approved companies eligible to receive advanced computing items including AI chips and servers without individual licenses; Commerce tied the change to the bilateral AI cooperation framework and said sensitive technology remains subject to anti-diversion commitments, so this is broader access for approved entities rather than unrestricted access. In the other direction, Reuters-syndicated reporting on July 8 said China is planning limited H200 access for top AI firms including Alibaba, ByteDance and DeepSeek, with the total possibly fewer than 200,000 chips. Separately, China imposed a temporary export ban on helium on July 10, a materials risk for fabs rather than a chip-access issue.
Latest coverage
Showing newest first with full-topic pagination
Rebellions said on July 23 that it ran SK Telecom's A.X K1 model on one RebelServer system and demonstrated simultaneous map-agent requests using Korean-built NPU infrastructure. UPI's August 7 coverage framed the result against a training market still dominated by Nvidia, citing Counterpoint data that 92% of surveyed sovereign language models used Nvidia chips.
Mobilint launched the MLD-R1 USB edge AI accelerator on Aug. 7, offering up to 10 TOPS of INT8 inference performance in a roughly 3 W device. Digital Today reports that the REGULUS-based accelerator connects through USB Type-C and supports Windows and Linux hosts. The product targets local inference workloads including computer vision, video analysis, robotics, security, and small language models.
Software researcher Daniel Lemire said on August 5 that the jump in memory costs had erased roughly 20 years of price declines. A Stanford-maintained retail dataset supports the direction of that claim: its July snapshot put the cheapest listed DDR5 memory at about $11.41 per GB, a level comparable with nominal DDR2 prices from the late 2000s, though the comparison depends on product generation and inflation treatment.
Anthropic is recruiting a custom-silicon team for Claude, with a Silicon Engineer listing updated August 4, 2026 and a related chip-design reinforcement-learning role updated August 3. The jobs span architecture through packaging and machine-learning-guided chip optimization, but Anthropic has not announced a processor specification, manufacturing partner, or production schedule.
AMD announced on August 6 that it reached a definitive agreement to acquire Toronto-based AI inference startup Taalas. Financial terms were not disclosed, and the transaction remains subject to customary closing conditions and regulatory approvals. AMD plans to integrate Taalas's model-specific silicon technology into its accelerator roadmap and pair it with Instinct GPU systems.
AMD, Spectro Cloud, and Supermicro announced AMD Instinct Coder on August 5, a turnkey on-premises and hybrid inference system for enterprise AI coding. The configuration pairs eight AMD MI325X GPUs in a Supermicro server with Spectro Cloud's policy-based model routing, directing suitable requests to local GLM-5.2 inference while retaining access to frontier models. The partners claim up to 70% token-cost savings.
Tom's Hardware reported on August 6 on a Hong Kong-based eBay listing for modified Nvidia RTX 2080 Ti cards with 22GB of VRAM at $499. The listing itself was last updated May 20, so the current event is renewed coverage rather than a new product launch. It offers used blower-style cards from unspecified board partners for local AI workloads.
Tamil Nadu's revised 2026-27 budget proposes Arivagam, an AI and innovation city, and allocates Rs 5 crore for a feasibility study. Finance Minister N. Marie Wilson said the planned city would combine an AI university, international skills center, semiconductor hardware testing lab, multi-domain incubators and residential infrastructure. No construction timetable, operating budget or site was disclosed in the retrieved reports.
Wolfspeed and LITEON announced on August 6 that Wolfspeed's silicon carbide MOSFET technology has qualified for LITEON's 800 VDC power sidecar and compute-rack power-supply platforms. The companies plan to pursue adoption across cloud-service-provider platforms and future hyperscale deployments, aiming for more efficient, dense and reliable power delivery while reducing component and procurement complexity.
South Korean NPU startups Rebellions, FuriosaAI and DeepX are turning demonstrations, deployments and fundraising into operating evidence ahead of possible public listings. Rebellions says one server ran SK Telecom’s 500-billion-parameter A.X K1 model, FuriosaAI is installing RNGD systems at an Equinix facility in Lisbon, and Dealsite reported a 2.85 trillion won pre-money valuation for DeepX’s Series D—progress that still falls short of comparable revenue proof.
SpaceX said on August 4 that it is working with Nvidia to design the compute payload for its Starmind AI1 satellite, with Rubin GPUs and Vera CPUs planned for each spacecraft. On the same day's earnings call, Elon Musk said SpaceX would build its broader AI infrastructure exclusively on Nvidia and projected up to 10 GW of compute by the end of 2027, a forward-looking target rather than deployed orbital capacity.
FuriosaAI said on August 4 that it will supply 1,800 RNGD inference accelerators for the first phase of a Stockholm AI data center being developed with I/ONX HPC and Velox. The first 2 MW of compute is expected online in early 2027, with another 8 MW later that year and more than 7,000 additional RNGD accelerators planned as the facility scales toward 15 MW.
Elon Musk said on August 4 that SpaceX had committed to using NVIDIA GPUs exclusively, while Wccftech reported that he identified Vera Rubin as the platform for the company's AI data centers. Wccftech also reported company targets of 2 gigawatts of compute by year-end and 10 gigawatts by the end of 2027; those remain forward-looking projections.
Samsung Electronics introduced zHBM and zNAND-O 3D-memory concepts and a more-than-400-layer V10 BV-NAND prototype on August 4 at FMS 2026 in Santa Clara. Samsung says a next-generation zHBM interface could deliver about eight times HBM5 performance, while V10 BV-NAND raises memory density by approximately 58% over V9.
Sandisk and SK hynix released the first Open Compute Project technical specification for High Bandwidth Flash on August 3-4. The open framework covers the xPU-HBF host interface, electrical design, packaging, reliability and software operations. SK hynix says the outlined configurations reach 512 GB and span about 0.4 to 3.0 TB/s, but the release is a standard rather than a commercial product launch.
Banana Pi launched the BPI-AI2N system-on-module and carrier board on March 19, 2025, pairing Renesas' RZ/V2N Vision AI MPU with dual camera inputs, dual Gigabit Ethernet and expandable storage. The company rates the accelerator at 15 sparse TOPS and documents Yocto and Armbian support, but it has not published model-level benchmark results.
Kioxia announced the GP1 Series PCIe 6.0 NVMe SSD on August 3, claiming up to 10 million random-read IOPS at a 512-byte block size using second-generation XL-FLASH. Evaluation samples are due to select customers by the end of 2026, so the headline performance remains a vendor specification rather than an independently validated production result.
South Korean officials and a Samsung SDS-led consortium broke ground on the National AI Computing Center in Haenam County on August 3, 2026. The public-private project targets completion of its initial build by 2028 with 15,000 advanced AI accelerators, creating a national compute hub for large-model training and research.
Doosan agreed on July 31 to buy SK Inc.'s 70.6% stake in silicon-wafer maker SK Siltron for 2.3 trillion won ($1.61 billion). The deal would add a major front-end chip material to Doosan's copper-clad-laminate and semiconductor-testing businesses, although closing and integration work still remain.
South Korean companies and U.S. technology firms announced AI semiconductor cooperation projects valued by Seoul at a reported $950 billion during a July 24 San Francisco summit hosted by President Lee Jae Myung. The package combines $750 billion in SK Group supply and infrastructure agreements with a Samsung-Broadcom memorandum of understanding estimated at more than $200 billion over five years, so the headline total represents planned multi-year cooperation rather than a single completed purchase.
Fubon Research projects that Google could deploy 12 million to 15 million ninth-generation TPUs in 2028, according to reports by Tom's Hardware and TechSpot. The analyst estimate would put Google's annual custom-chip volume near Fubon's forecast of 12.4 million Nvidia data center GPUs, but Google has not confirmed the target or disclosed TPU v9 specifications.
Amazon reported July 30 that AWS second-quarter sales rose 37% to $42.2 billion, its fastest growth in 18 quarters. CEO Andy Jassy told investors that Amazon now expects about $220 billion in 2026 cash capital spending but still will not have enough capacity to meet all demand this year, with the constraint expected to continue in 2027.
Samsung Electronics and SK Hynix posted a combined 150 trillion won in operating profit for the quarter ended June 30, as AI-server demand and higher memory prices lifted both companies to records. Samsung reported 89.5 trillion won on July 30, while SK Hynix reported 60.5 trillion won a day earlier. Both said supply remains tight and described multiyear customer agreements, extending the AI buildout's effects beyond GPU demand.
A Bloomberg report republished by The Japan Times on July 31 found that China's first-half growth was increasingly concentrated in technology hubs tied to AI hardware, while older industrial cities lagged. It contrasted memory-chip center Hefei with auto-manufacturing hub Changchun and cited Nomura analysis of regional output. National data separately showed high-tech manufacturing grew 13.3% in the first half, versus 1.3% growth in consumer-goods retail sales.
China's National Bureau of Statistics reported that integrated-circuit manufacturing profits rose 2,579.5% year over year in the first half of 2026, while profits across electronics industries increased 96.9%. The agency linked the broader electronics gain to rising computing demand as AI spread across sectors; separate NBS data reported by the South China Morning Post put integrated-circuit output near 280 billion units, up 23%.
The EuroHPC Joint Undertaking opened a tender on July 30 to select consortia for up to seven AI Gigafactories, with proposals due November 12. The official release says public anchor funding is intended to unlock more than €20 billion in private investment; the Associated Press reports the EU and member states are offering €10 billion in public financing.
SK hynix reported preliminary second-quarter revenue of 79.3 trillion won and operating profit of 60.5 trillion won on July 29, both quarterly records, with a 76% operating margin. The memory maker attributed the increase to strong AI-infrastructure demand and higher DRAM and NAND prices, and said HBM4 mass shipments began during the quarter.
Lava's new research identifies 36,872 internet-exposed baseboard management controllers, the processors that hold power, firmware and remote console control over enterprise and cloud servers. Nearly 25,000 leak password-derived material before login, and HPE iLO factory passwords fall to GPU cracking in under a minute. In an email interview with Lets Data Science, Lava said roughly 60 newly exposed endpoints appear every day, and that all 50 endpoints it found already compromised were HPE iLO 4 systems, carrying ransom notes from two separate campaigns.
Nvidia and Safe Superintelligence Inc. announced a long-term partnership on July 27 that will give Ilya Sutskever's lab access to Vera Rubin systems and increase its compute capacity tenfold. Nvidia disclosed an investment but not its size; Reuters reported a $5 billion equity investment, citing a person briefed on the deal.
Singapore startup Acrab announced its GELIX 1 edge AI system-on-chip and Agent Box reference system on July 23, targeting local inference for models in the 100-billion-parameter class. Acrab says the 5 nm chip combines a 20-core Arm CPU and multicore NPU with 273 GB/s unified-memory bandwidth; independent reports describe its 7.5x M4 Pro result as an unverified company benchmark.
