Semiconductor Watch: AMD Helios And Broadcom's Jalapeño Chip Enter Production Stage, Samsung and AMD Sign AI Memory MoU
The Chip Shifts That Actually Moved the AI Supply Chain Today
AMD Helios Enters Production With Samsung HBM4 as Nvidia Raises HBM4 Specs to Stay Ahead
AMD's rack-scale push went into production overnight. The Korea Herald confirmed Aug 18 that AMD has begun production of Helios, its AI server rack positioned against Nvidia's NVL72 systems, with Samsung Electronics supplying high-bandwidth memory inside. AMD Chair and CEO Lisa Su held up the Instinct MI455X accelerator at Advancing AI 2026 in San Francisco Thursday. According to Chosun Biz and TrendForce, MI450 is set to feature 432GB HBM4 per GPU — about 150GB more than Nvidia Vera Rubin — while Samsung passed final quality validation for Nvidia Rubin HBM4 and is set to improve yield from ~50% in Q4 2025 to 60-70% in H2 2026.
Competitive pressure is pushing specs higher. TrendForce reports Nvidia has pressed key suppliers of Vera Rubin racks to upgrade HBM4 speed per pin to 10 Gbps and boosted bandwidth 10% to 22.2 TB/sec via faster 8-Hi stacks to surpass MI455X. TweakTown confirmed Nvidia updated Rubin NVL72 at CES 2026 to 22.2 TB/sec. SK hynix is expected to maintain a 60-70% share in the early mass-production phase per TrendForce; Samsung 25-30%; Micron is supplementary, chasing 30% by year-end. HBM4 sold out for 2026 as of January.
Key facts: AMD Helios in production with Samsung HBM4 · MI450 432GB HBM4 vs Rubin ~282GB · Nvidia Rubin bandwidth +10% to 22.2 TB/sec · HBM4 pin speed 10 Gbps · SK hynix 60-70% share early · Samsung yield 50% → 60-70% H2 2026.
Why it matters: Rack-scale architecture makes HBM capacity and bandwidth the binding constraint, not just GPU count. Who is affected: Samsung, SK hynix, Micron gain pricing power; Astera Labs, Broadcom networking benefit from scale-up fabric. What to watch next: Whether Samsung hits 70% output ramp and whether Nvidia holds 10 Gbps spec through qualification.
Source: Korea Herald — Samsung supplies HBM4 for Helios; TrendForce — Nvidia raises HBM4 specs; TweakTown — Rubin 22.2 TB/sec.
Samsung and AMD Sign AI Memory MoU as HBM Shortage Extends to 2028
Samsung Electronics and AMD signed an MoU on AI memory this week during Nvidia's GTC week, where Jensen Huang announced a foundry partnership with a Korean firm and praised its HBM4 chips, per Reuters. The tie-up highlights the race to lock in long-term supply partnerships for advanced memory as AI demand tightens supply. Last month AMD said it agreed to sell up to $60B worth of AI chips to Meta Platforms, making memory security critical. Samsung, which counts Nvidia and AMD among HBM customers, said it expects HBM4 revenue to more than triple in Q3 and sees the chip shortage extending to 2028, signing long-term supply deals with key customers.
The MoU explores a foundry partnership beyond memory, signaling broader Korea-US supply chain alignment. Micron meanwhile disclosed an HBM4 product launch in 2026, followed by HBM4E likely for Nvidia Rubin R100 and AMD successor to Instinct MI400x per TechRadar, aiming for a bigger slice of the $100B HBM market. Nvidia's Rubin Ultra was originally unveiled with 1TB HBM4E per GPU, though early samples use HBM4. Market now in a seller's market, with multi-year lock-ins replacing spot buying.
Key facts: Samsung-AMD MoU on AI memory + foundry · AMD $60B Meta deal last month · Samsung HBM4 revenue to more than triple Q3 · Shortage seen extending to 2028 · $100B HBM market · Micron HBM4 2026 + HBM4E for Rubin R100.
Why it matters: Long-term supply agreements replace transactional buying — memory becomes a strategic asset like GPUs. Who is affected: Cloud providers securing multi-year HBM allocations; consumer DRAM/GDDR squeezed as 70-80% capacity shifted to HBM/server DRAM. What to watch next: Samsung Q3 HBM4 output ramp +70% to meet Nvidia/AMD demand per Ked Global.
Source: Reuters — Samsung and AMD sign MoU; Ked Global — Samsung ramps HBM4 70%; TechRadar — Micron HBM4/HBM4E 2026.
Broadcom's Jalapeño Chip With OpenAI Moves Toward Production as Custom Silicon Race Intensifies
OpenAI's first custom AI chip is moving from samples to deployment. On its Q2 fiscal 2026 earnings call, Broadcom disclosed silicon for OpenAI had already been delivered and remained on track for production, with initial deployment planned by the end of 2026. The chip, called Jalapeño, unveiled June 24, was built from scratch for LLM inference with the goal of improving efficiency and lowering cost to serve AI systems at scale, with companies showing ~50% cost savings vs GPUs. Broadcom CEO Hock Tan said the chip is as good as Blackwell chips or Google TPUs, per Reuters, and expects AI semiconductor revenues to exceed $100B in 202,7 supported by OpenAI and other hyperscale customers.
The table as of July shows where custom programs stand: Meta Iris MTIA with Broadcom TSMC production starts September 2026, OpenAI Jalapeño with Broadcom TSMC unveiled June 2026 deployment year-end, Google TPU Ironwood 7th gen with Broadcom TSMC generally available late 2025, Amazon Trainium 3 with Marvell TSMC generally available since Dec 2025 per TechTimes. Broadcom now designs chips for three rivals — Meta, OpenAI, Google — making it a central bottleneck for custom silicon. ASIC designers like Broadcom and MediaTek will have a working HBM4 solution in 2027 chips, but Nvidia and AMD will have custom HBM4 in 2026 products, per TechPowerUp.
Key facts: Jalapeño silicon delivered · Production deployment by end of 2026 · ~50% cost savings vs GPUs · Broadcom AI semi revenue >$100B expected in 2027 · Broadcom designs for Meta, OpenAI, Google · Multi-generation platform with Celestica racks.
Why it matters: Custom silicon shifts value from GPU FLOPs to inference efficiency and portability, a direct challenge to CUDA's moat. Who is affected: Qualcomm's $3.92B acquisition of Modular for Mojo/MAX portable stack gains relevance; Nvidia faces first credible inference chip alternative. What to watch next: Whether Jalapeño meets year-end deployment and whether cost savings hold at scale.
Source: Zacks — Broadcom OpenAI Jalapeño upside; TechTimes — Custom silicon table July 2026; Reuters — OpenAI unveils custom chip.
Onsemi and Analog Devices Point to AI Power as Fastest-Growing Business
Analog and power vendors are confirming AI as a synchronized growth driver. Onsemi reported Q2 2026 results Aug 18, with CEO Hassane El-Khoury stating,g "AI data center remains our fastest-growing business, and we now expect revenue to more than double in 2026, demonstrating strength of our intelligent power portfolio and growing customer adoption across power tree." CFO Thad Trent said year-over-year EPS grew four times faster than revenue,ue driven by gross margin expansion. On Monday, onsemi forecast Q3 revenue above Wall Street expectations, betting on surging demand for power management chips used in AI data centers, as it looks to a $7B all-stock acquisition of Synaptics announced in June to tap AI devices and robotics.
Analog Devices projected to report $3.92B revenue for fiscal Q3,3 up 36% Y, oY with adjusted EPS $3.34 vs $2.05 a year ago, per Visible Alpha, with UBS seeing the white-hot data center industry boosting revenue above seasonal trends for at least the next 4 quarters, per Investopedia. Texas Instruments Q1 analog revenue +22% YoY to $3.92B, with data-center +90% YoY to a >$1B run rate and Q2 guide $5.0-5.4B vs $4.86B consensus. TI surged ~20% after earnings, lifting ADI and ON to record highs per Barron's. All three are part of Nvidia's power partner ecosystem for GB200/GB300 NVL72.
Key facts: Onsemi AI data center revenue to more than double in 2026 · ADI Q3 projected $3.92B, +36% YoY · TI analog +22% YoY · TI data-center +90% YoY >$1B · Synaptics $7B all-stock deal.
Why it matters: Analog and power shift from cyclical auto/industrial laggards to AI infrastructure growth — power tree determines rack density. Who is affected: Texas Instruments, ADI, onsemi, Infineon, Renesas benefit from 800V DC and high-density racks. What to watch next: Whether onsemi doubles again in H2 and whether Empower acquisition by ADI for $1.5B delivers 4x footprint shrink.
Source: GlobeNewswire — onsemi Q2 2026; Reuters — Onsemi Q3 forecast; Investopedia — ADI Q3 preview $3.92B +36%.
Infineon Forecasts Strong Revenue Growth as Price Hikes Spread Across Power and Analog
Infineon Technologies said it expects strong revenuegrowth inn thcurrenttt fiscal year as the race to build AI infrastructure keeps adding fuel to red-hot semiconductor demand, per Morningstar, Aug 18. The company boosted its FY2026 AI power revenue target to €1.5B from €1B a quarter earlier and sees €2.5B in FY2027 +66% YoY, raising capex from €500M to €2.7B for Dresden capacity for data-center power supplies. Infineon, Analog Devices, and NXP Semiconductors have all announced price increases in 2026 driven by AI demand, inflation, and rising manufacturing costs, per Microchip USA, impacting power semiconductors, analog devices, and critical components across multiple industries.
The price hikes mark a regime change — analog vendors that historically fought deflation now have pricing power tied to AI rack density. TI holds >10x share vs 10th-ranked Renesas in analog, per company reports. Grid modernization and 800V DC buses drive demand beyond data centers into energy infrastructure. CEO Jochen Hanebeck said grid modernization and 800V DC buses are driving demand. Zacks Aug 14 highlighted TI, ADI, and ON as three top-ranked names enjoying a synchronized global recovery the market hasn't fully priced in.
Key facts: Infineon AI power €1.5B FY26 → €2.5B FY27 · Capex +€500M to €2.7B Dresden · TI, ADI, NXP, Infineon price increases 2026 · TI >10x share vs Renesas 10th · Zacks top-ranked TI, ADI, ON.
Why it matters: Analog/power pricing power indicates shortage, not just growth — 800V DC and high-density AI racks create a new power tree bill of materials. Who is affected: Server OEMs face rising BOM; power vendors gain margin expansion. What to watch next: Whether price increases stick through 2026 and whether Dresden capacity comes online in time.
Source: Morningstar — Infineon forecasts strong growth on AI demand; Microchip USA — Infineon, ADI, NXP price increases 2026; Zacks — TI, ADI, ON top ranked.
The AI Factory Moves From Chip to Rack-Scale System
The unit of compute is now the rack, not the GPU. AMD Helios with 72x MI455X GPUs, 432GB HBM4 per accelerator, 19.6 TB/s bandwidth plus 800GbE Pensando networking shipping since July competes directly with Nvidia GB300 NVL72 with 72 Blackwell GPUs via NVLink now in production at AWS, Google and CoreWeave. Both require a 7MW power/cooling blueprint co-developed by Vertiv, with Eaton acquiring Boyd Thermal for thermal management, cooling lead times now 5x longer than in 2023, and cold-plate shipments projected from 8M in 2025 to 356M by 2030.
The shift redefines semiconductor TAM. Data centers now consume 70% of memory per TrendForce, with 70-80% of new DRAM capacity toward HBM/server DRAM, while 800G/1.6T optical modules and high-speed switch chips are in global shortage. AI server shipments forecast +31% YoY in 2026 per TrendForce. Top six hyperscalers: $700B capex in 2026 per Moody's vs $615B March forecast; Big Five: $600B+ +36% YoY, with 75% or $450B directly tied to AI infrastructure. Rack-scale systems make networking, power, and cooling as strategic as accelerators.
Key facts: Helios 72x MI455X 432GB HBM4 19.6 TB/s · GB300 NVL72 shipping · Vertiv 7MW blueprint · Cooling lead times 5x · Cold-plate 8M → 356M 2030 · AI servers +31% YoY · $700B top six capex.
Why it matters: The semiconductor market is now defined by AI factories, not just accelerators — memory, networking, analog/power, and cooling become part of the semiconductor narrative. Who is affected: Astera Labs, Broadcom, Marvell, Vertiv, Eaton, and Schneider benefit. What to watch next: Whether rack-scale becomes standard procurement unit and whether power equipment delays constrain shipments despite chip supply.
Source: AMD Helios; Nvidia GB300 NVL72; TrendForce AI servers +31%.
Reviewed by Erwin Castro
on
Wednesday, August 19, 2026
Rating: