Daily Update - August 4th, 2026
SK Hynix and SanDisk publish the first OCP High Bandwidth Flash spec, Tower guides to a record $520M and lifts its 2028 model to $3.6B, onsemi expects AI data center revenue to more than double.
Good morning, it’s Tuesday, August 4.
SK Hynix and SanDisk dropped the first OCP High Bandwidth Flash spec. Tower and onsemi both posted record AI-driven quarters, and Tower raised its 2028 revenue target by $800 million.
Let’s get into it. — Austin & Vik
Be sure to check out the Semi Doped podcast on YouTube or your favorite podcast player!
SK Hynix and SanDisk Release First OCP High Bandwidth Flash Spec at FMS 2026
SK Hynix and SanDisk published the first OCP technical specification for High Bandwidth Flash on August 3, 2026, formally establishing HBF as a new memory-storage tier targeting AI inference workloads. The spec was developed through the Open Compute Project, with Google and Tenstorrent backing the effort, giving the standard immediate credibility across both hyperscaler and AI accelerator communities. FMS 2026 ran August 4-6 at the Santa Clara Convention Center, making it the public debut stage for a format the two companies have spent considerable engineering effort positioning between DRAM and conventional NAND.
The specification’s release through OCP provides the formal foundation that third-party silicon vendors need to build HBF-compatible controllers and interconnects. TrendForce reported the standard is aimed directly at AI memory bottlenecks that have grown acute as inference deployments scale. Kioxia used the same conference floor to press its own case in the tier, announcing the GP1 Series SSDs built on XL-FLASH Gen 2 reaching 10 million random read IOPS and separately showing PCIe 6.0 NVMe prototypes. TrendForce noted Kioxia is also eyeing PCIe 6.0 and UFS 5.0 mass production as it works to close ground on its South Korean rivals.
On an adjacent front:
Marvell announced it would showcase AI memory and storage infrastructure for agentic inference at FMS 2026, adding its silicon portfolio to the week’s converging narrative.
Credo Technology exhibited interconnect products for AI memory and storage applications at the show, targeting the connectivity layer that HBF and high-IOPS SSDs both require.
Silicon Motion presented next-generation NAND flash controllers for agentic AI and physical AI applications, spanning workloads from AI factories down to edge deployments.
Vik: HBF seems to have reached some sort of checkpoint at FMS 2026. I wrote about the implications of HBF in September 2025. Our memory problem now is so acute that HBF may workout as a viable tier. It has its downsides, but not relying on DRAM supply has its upsides.
Austin: Spec opens door for second-source competition. On the stack side, Samsung has the motive (behind in HBM), YMTC has the bonding tech plus a domestic market that isn’t permitted to buy HBM, and Micron has everything it needs… but maybe would rather aim its TSV capacity at HBM4?
The blocker isn’t NAND supply though, it might be CoWoS supply as HBF sits on the same interposer as the accelerator, so a second needs packaging allocation too… or maybe nobody has to. Intel’s EMIB-T runs power through the bridge which is exactly what an HBM4-class stack needs. SK hynix is already reported to be testing EMIB with its own HBM, and Google, one of the two names backing the spec, has reportedly booked Intel to package TPUs. So HBF could be built at Intel Foundry…
Tower Lifts Its 2028 Model by $800 Million as onsemi Guides AI Data Center Revenue to Double
Tower Semiconductor reported record second quarter revenue of $460 million, up 24% year over year, and guided third quarter revenue to $520 million, a 31% year-over-year increase that would set another company record. Silicon photonics carried the quarter: SiPho revenue reached a $680 million annual run rate, up from $180 million a year earlier, with Tower expecting to cross $1 billion in run rate during the fourth quarter. Net profit excluding non-recurring items was $91 million, a 20% net margin and 95% above the $47 million posted a year ago.
The more consequential number is the long-range model. In May, Tower told investors it was targeting $2.8 billion of revenue and $750 million of net profit in 2028. Three months later CEO Russell Ellwanger raised that to “$3.6 billion of revenue with $1.2 billion net profit, being fully spoken for by our customers.” That is an $800 million revenue raise and a 60% increase in the profit target inside a single quarter, on capacity the company says is already committed.
onsemi described the same demand from the power side a day earlier. Revenue of $1,604 million grew 9% year over year and 6% sequentially, non-GAAP gross margin reached 39.3%, and free cash flow of $425.4 million quadrupled from a year ago. CEO Hassane El-Khoury said “AI data center remains our fastest-growing business, and we now expect revenue to more than double in 2026.” Third quarter revenue is guided to $1.65 billion to $1.75 billion, and the company pointed to an expanded role in the Nvidia MGX ecosystem, EliteSiC and MOSFET wins with China cloud power supplier Great Wall, and the launch of GaNEXUS, its 40V to 650V gallium nitride portfolio.
Austin: AI continues to drive optics and power markets, benefitting the specialty-analog and foundry folks.
Sector Watch
Memory
HP, Asus and Acer begin using CXMT DRAM chips in consumer PCs amid ongoing memory shortage, according to Nikkei Asia. (Nikkei Asia)
Commodity DRAM and NAND flash prices hit fresh record highs in July, extending a supply-tight rally per The Elec market data. (The Elec)
Montage Technology began trial production of a CXL 3.2 memory expander controller targeting AI data center memory expansion.
Optics & Networking
ZJ Innolight debuts on the Hong Kong Stock Exchange in a record-setting IPO as indium phosphide substrate shortages tighten supply for AI optical modules. (LEDinside)
TOHKnet and Adtran complete Japan’s first 50G PON trial on a live network, advancing next-generation access speeds for carrier deployments. (Adtran)
Philippine telcos PLDT, Globe and Converge jointly propose a $500 million domestic subsea cable system to the government, pooling resources to cut foreign transit dependence. (Light Reading)
Compute & Edge
Silicon Labs launches the BG2B, its lowest-power Bluetooth LE SoC, integrating Channel Sounding, Secure Vault security and CAN-FD peripherals for edge IoT. (Silicon Labs)
Rebellions begins shipping ATOM-Max NPU servers this month to AI CCTV control centers in South Gyeongsang Province and Ulsan, South Korea. (The Elec)
d-Matrix acquired Wallaroo.ai to build an end-to-end AI inference platform orchestrating heterogeneous chips, expanding beyond its in-memory compute hardware. (PR Newswire)
Data Centers
CoreWeave and Leidos partner to deliver an AI compute offering specifically targeting US intelligence community workloads, per Data Center Dynamics. (Data Center Dynamics)
NTT Data is eyeing a $9 billion capital outlay for data center expansion across Japan as domestic AI infrastructure demand accelerates. (Bloomberg.com)
Wistron contributes 2.4 million computing hours to Taiwan’s sovereign AI initiative, backing the government-led effort to build domestic AI capacity. (digitimes)
Policy & Trade
China revises chip-design protection regulations to strengthen IP rights and spur domestic semiconductor innovation amid sustained US export controls. (SCMP)
India seeks parliamentary approval for sweeping tax cuts targeting foreign electronics manufacturers and global funds to accelerate local chip and device production. (Bloomberg Tech)
Marvell Technology will hold its Q2 FY2027 earnings call and an Investor Day on October 6, 2026. (Marvell)