Semiconductor Watch: US Tariff Policy, Nvidia–MediaTek Deepens, Samsung's CUBE Memory Strategy & Broadcom's Blowout Quarter
U.S. Semiconductor Tariffs Take Shape as Broadcom Stuns with $350B AI Revenue Forecast
Tariffs, Trillions, and the New Semiconductor Order
On September 2, U.S. Commerce Secretary Howard Lutnick confirmed the Trump administration is crafting a "targeted" and "thoughtful" tariff policy for semiconductors—tying tariff relief directly to domestic manufacturing investment. The policy arrives alongside stunning financial results from Broadcom, which reported Q3 2026 revenue of $295.91 billion (up 86% YoY) and projected AI semiconductor revenue will double to approximately $1.15 trillion in fiscal 2027 and reach $2.3 trillion by fiscal 2028.
Read together, these signals point to a semiconductor industry at an inflection point: governments are weaponizing tariffs to reshape manufacturing geography, while AI demand is producing financial results that defy historical comparison. This edition also covers ASML's High-NA EUV entering mass production, Samsung's CUBE memory strategy targeting HBM5 at double HBM4E performance, and the deepening Nvidia–MediaTek partnership.
Market Signal
U.S. Semiconductor Tariffs Take Shape as Broadcom Stuns with $350B AI Revenue Forecast
On September 2, U.S. Commerce Secretary Howard Lutnick confirmed the Trump administration is crafting a "targeted" and "thoughtful" tariff policy for semiconductors—tying tariff relief directly to domestic manufacturing investment. The policy, which Lutnick described as "if you build here, you don't pay, but if you don't build here, expect to pay to enter the greatest market in the world," arrives alongside stunning financial results from Broadcom. The networking and ASIC giant reported Q3 2026 revenue of $295.91 billion, up 86% year over year, with CEO Hock Tan projecting AI semiconductor revenue will double to approximately $1.15 trillion in fiscal 2027 and reach $2.3 trillion by fiscal 2028. Read together, these signals point to a semiconductor industry at an inflection point: governments are weaponizing tariffs to reshape manufacturing geography, while AI demand is producing financial results that defy historical comparison.
What Changed
U.S. Tariff Policy. Commerce Secretary Lutnick confirmed the administration is working on semiconductor-specific tariffs that would exempt companies manufacturing in the U.S. while imposing duties on imports. The policy could dramatically expand the number of tech products subject to tariffs, affecting not just chips but laptops, gaming consoles, and data center servers. Country-specific tariff rates and quotas are under consideration. Lutnick cited TSMC's $265 billion Arizona investment and Micron's $250 billion U.S. memory fab as evidence the policy is already working, noting the administration has secured $1.2 trillion in U.S. semiconductor investment commitments. He added that Taiwan is expected to announce an additional $20–30 billion in U.S. investment next week.
Broadcom's Blowout Quarter. The company reported Q3 2026 revenue of $295.91 billion, up 86% year over year, with adjusted EPS of $3.32, up 96%. CEO Hock Tan raised 2026 AI revenue guidance from $56 billion to $58 billion and projected AI semiconductor revenue will double to ~$1.15 trillion in 2027 and double again to $2.3 trillion in 2028. The company has begun mass production of Google's next-generation TPU 8I, which Tan said matches or exceeds Nvidia's Vera Rubin performance. Broadcom also confirmed it has secured supply chain commitments through 2027 to support this growth. The stock initially fell over 6% after hours before reversing gains during the earnings call.
Nvidia–MediaTek Deepens. MediaTek CEO Rick Tsai confirmed Nvidia's $3.5 billion investment is "a very strategic investment" that demonstrates confidence in both MediaTek and Taiwan. Alphabet also participated in the bond offering, with Tsai noting the strategic significance of having both ecosystem partners invest. MediaTek maintains its Q4 2026 AI ASIC revenue target of over $2 billion. The partnership extends NVLink Fusion to custom XPUs, positioning Nvidia's interconnect as the substrate for custom AI chips. TSMC's capacity remains sufficient for MediaTek's needs, Tsai said—"TSMC, no problem!"
Samsung's CUBE Memory Strategy. At SEMICON Taiwan 2026, Samsung unveiled its next-generation memory strategy "CUBE"—standing for Capacity, Utilization, Bandwidth, and Efficiency. The company set development targets for HBM5 at approximately double the performance of HBM4E, with a 20% improvement in performance-per-watt and a 20% reduction in thermal resistance. Samsung plans to use its 2nm foundry process for the base die, with mass production targeted for around 2028. The strategy extends vertical integration beyond HBM to NAND flash with "zNAND-O" for on-device and edge AI.
ASML High-NA EUV Enters Mass Production. At SEMICON Taiwan 2026, ASML confirmed its High-NA EUV lithography systems have entered large-scale production. Global High-NA EUV systems have now exposed over 1.35 million wafers, with 10 systems operational across four customers (Intel, Samsung, SK hynix, and TSMC). The EXE:5200B system has achieved 135 wafers per hour throughput—35% above the previous generation—with fleet availability at 84% and is expected to reach 90% by Q4 2026. Intel's 18A process is the first to use High-NA EUV for mass production of its Panther Lake Core Ultra Series 3 processors. ASML CEO Christophe Fouquet confirmed: "Intel is now basically using High-NA for production on its most advanced products. This is a very important milestone that proves the maturity of the equipment".
Navitas Targets First U.S.-Made GaN Chips. Navitas Semiconductor targets first wafer shipments in September of its Gen 5 GaNFast chips manufactured with GlobalFoundries in the U.S., marking a milestone for domestic compound semiconductor production.
Strategic Analysis
1. The tariff threat is becoming a reality—with a "build in America" carve-out. Lutnick's confirmation that semiconductor tariffs are being actively developed—with country-specific rates and quotas under consideration—represents a significant escalation from previous rhetoric. The policy explicitly ties tariff relief to U.S. manufacturing investment. For TSMC, Samsung, and SK hynix, this creates a clear calculus: invest in U.S. capacity or face escalating duties on products entering the world's largest semiconductor market. Lutnick's claim of $1.2 trillion in U.S. semiconductor investment commitments suggests the policy is already influencing corporate decisions. The inclusion of finished goods like laptops and servers in tariff considerations means the policy could affect downstream electronics manufacturers as well.
2. Broadcom's numbers defy comprehension—and signal ASIC's ascendancy. $2.3 trillion in projected AI semiconductor revenue by 2028 is not a forecast; it is a declaration that AI silicon is becoming the largest category in the history of the semiconductor industry. Broadcom's dominance in custom ASICs (Google's TPU, Meta's inference chips) positions it as the primary alternative to Nvidia's GPU-centric model. That Broadcom's TPU 8I matches or exceeds Nvidia's Vera Rubin performance suggests the custom silicon market is maturing faster than expected. The company's supply chain commitments through 2027 indicate Broadcom is securing capacity years in advance—a defensive move against the component scarcity that has plagued the industry.
3. Nvidia's MediaTek bet is about ecosystem lock-in, not just investment. MediaTek CEO Rick Tsai's comments about the "very strategic" nature of Nvidia's investment, alongside Alphabet's participation, reinforce the interpretation that Nvidia is building a coalition—not just writing a check. By embedding NVLink Fusion into MediaTek's custom XPU design capability, Nvidia ensures that even custom chips built by third parties remain tethered to Nvidia's rack-scale ecosystem. As one analyst noted, the competition is moving from "how many GPUs Nvidia sells" to "who defines the connection architecture of an AI factory". MediaTek's AI ASIC revenue target of over $2 billion in Q4 2026 suggests the partnership is already generating tangible revenue.
4. Samsung's CUBE strategy signals memory's shift from component to system. Samsung's CUBE announcement reflects a broader industry trend: memory is no longer a discrete component but an integrated part of the AI system design. By targeting HBM5 at double HBM4E performance with improved power efficiency and thermal management, Samsung is acknowledging that memory is becoming the bottleneck for AI system performance—and that solving it requires co-design with logic. The use of 2nm foundry for the base die underscores the increasing integration between foundry and memory capabilities. Mass production targeted for 2028—a timeline consistent with SK hynix's Indiana plant—suggests HBM5 will be the memory workhorse for the next generation of AI systems.
5. High-NA EUV's production milestone changes the lithography roadmap. Intel's use of High-NA EUV for mass production of Panther Lake processors validates the technology's readiness for leading-edge logic. The 135 wafers-per-hour throughput and 84% availability (rising to 90%) address the primary concerns about High-NA's productivity. For TSMC, Samsung, and SK hynix, this means High-NA EUV is no longer a future option but a present competitive necessity for the most advanced nodes.
6. The semiconductor equipment market continues to expand. SEMICON Taiwan 2026 and CSEAC 2026 showcased a broadening equipment ecosystem. Chinese equipment maker AMEC introduced an extreme high-aspect-ratio etcher with 70–90:1 capability, while Tongtai Group reported semiconductors now account for 40% of its machine tool and electronics equipment revenue. Hanwha Semitech unveiled four advanced packaging tools including FOPLP and hybrid bonding equipment. The equipment market is expanding beyond the traditional ASML/Applied Materials/Lam Research axis as both Chinese domestic players and new entrants from adjacent industries enter the space.
Enterprise & Industry Impact
- For semiconductor manufacturers: The U.S. tariff policy creates a clear "build in America" incentive. Companies must now weigh the cost of U.S. capacity investment against the risk of escalating tariffs on imports. TSMC's $265 billion Arizona investment and Micron's $250 billion U.S. memory fab suggest the largest players are already responding.
- For AI infrastructure buyers: Broadcom's $2.3 trillion 2028 AI revenue projection signals that AI semiconductor costs are not plateauing. Enterprises planning multi-year AI infrastructure budgets should expect continued pricing power on the silicon side, particularly for custom ASICs and networking components.
- For competitors (AMD, Intel, Nvidia): Broadcom's ASIC momentum—now matching Nvidia's Vera Rubin performance with Google's TPU 8I—challenges Nvidia's GPU-centric dominance. Meanwhile, Intel's High-NA EUV production milestone could restore its manufacturing competitiveness if the 18A process delivers.
- For memory buyers: Samsung's HBM5 targeting 2028 mass production aligns with SK hynix's Indiana timeline, confirming that HBM capacity constraints will persist through at least 2028. The 70% of new capacity allocated to HBM by the three major memory makers remains insufficient to close the 50–60% capacity gap.
- For policy watchers: The Trump administration's semiconductor tariff policy is moving from concept to reality. The inclusion of finished goods like laptops and servers means the policy could affect a much broader range of electronics imports than previously anticipated.
What to Watch Next
- Broadcom's post-earnings stock trajectory. The stock fell over 6% after hours before reversing during the call—volatility that reflects market uncertainty about whether AI semiconductor growth is sustainable or approaching a peak.
- TSMC's capacity expansion. With TSMC acknowledging it "cannot keep pace with the AI boom's growth trajectory" despite fivefold capacity expansion, any announcement of accelerated capacity investment would signal the company's confidence in sustained demand.
- Samsung's HBM5 customer engagements. The company's 2028 mass production timeline will depend on customer adoption. Any announced design wins would validate the CUBE strategy.
- The FCC's optical transceiver import rule. With the semiconductor tariff policy now moving forward, the previously drafted restrictions on Chinese optical transceivers may gain new momentum.
- MediaTek's Q4 2026 AI ASIC revenue. The company's target of over $2 billion will be the first concrete test of whether the Nvidia partnership is delivering tangible financial results.
- Intel's High-NA EUV yield data. As the first to use High-NA EUV in mass production, Intel's yield and cost data will determine whether the technology delivers on its promise or introduces new production challenges.
September 3, 2026, may be remembered as the day the semiconductor industry's scale became truly incomprehensible. Broadcom's projection of $2.3 trillion in AI semiconductor revenue by 2028 is not a forecast—it is a signal that the industry is entering a phase of growth without historical precedent. Meanwhile, the U.S. government is weaponizing tariffs to reshape where that manufacturing happens. High-NA EUV has entered mass production. Samsung is planning HBM5 at double HBM4E performance. Nvidia is building an ecosystem around custom silicon.
The semiconductor industry is no longer just a technology industry. It is an economic and geopolitical battleground where the stakes are measured in trillions of dollars, and the outcome will determine who controls the infrastructure of the AI era.
Sources
- Yonhap News Agency — Lutnick: Trump gov't considering 'targeted,' thoughtful' tariff policy for semiconductors
- Yonhap News Agency (Korean) — U.S. Commerce Secretary hints at 'targeted semiconductor tariffs'
- 21st Century Business Herald — Nvidia invests $3.5 billion in MediaTek, competing for 'rack-level' AI chip entry
- Liberty Times — MediaTek CEO Tsai: Nvidia investment shows confidence in Taiwan
- Dahe Cube — Broadcom CEO expects to deliver ~$350 billion in AI semiconductor products over next two years
- Seoul Finance — NYSE rebounds, Nvidia surges 3.2%
- Zaikei Shimbun — Samsung announces next-generation memory strategy 'CUBE', targets HBM5 at 2x HBM4E performance.
- Commercial Times — AMEC unveils multiple new etching and deposition tools
- Commercial Times — Tongtai Group targets semiconductor equipment market
- SECCW — ASML High-NA EUV enters mass production, Intel 18A process first to adopt
- Longbridge — TSMC edged higher, caught between strong AI demand and capacity constraints
- Eastmoney — UBS forecasts HBM demand to grow 90% in 2026, 77% in 2027
- FISCO — Nvidia rises on report of Hugging Face acquisition
The CODEW Stat
Broadcom projects AI semiconductor revenue will reach $2.3 trillion by fiscal 2028—a number larger than the GDP of most countries and a clear signal that the AI hardware buildout is entering a phase of growth without precedent in the history of the semiconductor industry.
Reviewed by Erwin Castro
on
Thursday, September 03, 2026
Rating: