Yesterday, Goldman Sachs and Citibank held TMT conferences simultaneously; here are some highlights from GIR. This session contains a large amount of information, and both conferences are still ongoing—we’ll continue bringing you the latest updates in real time.

Article author: Bei Xiang Mu Feng

Source: Wall Street Journal

TMTB: Astra also received significant attention over the weekend.

Although early consensus held that it has not yet delivered a comprehensive leap beyond Fable 5.1, aside from 3D modeling, spatial reasoning, and a few other use cases, more importantly, given the lineup and cadence of model releases prior to OAI’s DevDay on September 29 (Grok 4.7, Fable 5.1, META’s Watermelon, etc.), investors appear more confident about the path forward in September.

Meanwhile, we are seeing these models applied more specifically to workplace tasks via Grok Bot and consumer-facing Instinct, with META’s newly launched Muse also joining today.

The bullish interpretation in the semiconductor sector is: if better tools enable intelligence to operate across a broader range of tasks, you may not need a leap in intelligence itself—this could provide the elusive next major use case that semiconductor bulls have been waiting for since programming.

OpenAI: The overall tone is optimistic, in line with expectations

The core focus is enterprise business. For consumers: the company shifted its consumer-to-enterprise ratio from 60:40 at the start of the year to 50:50 by mid-year; in July, total annualized revenue increased by 20% month-over-month and 32% on an enterprise basis, despite a large base. The real new metric is usage intensity: the top 10% of enterprise customers now use tokens eight times more than average customers (up from three times previously), and internally at OpenAI, it’s 33 times. Friar defines this 33x ratio as a preview of where existing customers are headed—meaning enterprise growth isn’t driven by acquiring new clients, but by increasing consumption per customer. Codex growing from 100,000 to 25 million users, and Canva generating 100% of its code via OpenAI, both serve as supporting evidence for this trend.

Model: Focus on cost per task, not cost per token—requires 68% fewer output tokens than competitors for the same results. Most aggressive in responses to open-weight models: Luna deployed on Cloudflare is cheaper than GLM 5.3, meaning leading labs can bring inference costs below open-source self-hosting costs. Pricing strategy: subscription → usage-based → outcome-based; explicitly aims to move away from token-based billing toward vertical revenue sharing.

Compute: Still feels extremely scarce, with weekly trade-offs between training, research, and inference; full-stack is a spectrum, not a binary choice, and in-house building is increasing. ROI is measured by comparing the revenue generated so far for each model series plus the expected revenue over the next 12 months against the compute invested, to see if it turns significantly positive; gross margin expansion comes from rising revenue per gigawatt and declining cost curves. In response to criticism of overinvestment, she says this year has truly delivered returns.

AMD rose 6% on strong buying pressure driven by increased CPU demand across computing use cases. The CFO and CEO expressed confidence in the ramp progress of MI450/Helios during their Citigroup meeting, despite market concerns around execution. They noted that demand and shipment volumes for 2027 now exceed AMD’s original plans, with forecasts from three anchor customers—Meta, OpenAI, and Anthropic—all surpassing AMD’s expectations. The current constraint is supply, not demand.

Optical communications (LITE +11% / COHR +7%)

承接花旗 TMT 会议中对 AMD 的买盘：该公司计划在 2027 年下半年推出的加速器上，为更大规模的扩展域（可参考我们此前发布的一篇文章；昨日康宁宣布光纤订单已签约至 32 年，印证了需求确实旺盛：scale-across，即跨数据中心互联的 TAM）同时提供基于铜缆和近封装光互连的解决方案，铜缆与光通信很可能在多个产品世代中并存，而非经历一次全行业的单一切换。

This is the first time a commercial GPU vendor has provided a clear roadmap to bring optical communication into the intra-domain (where bandwidth per GPU is a multiple of that outside the scale-out domain), rather than stopping at the switch level. This validates comments made by LITE’s CEO last week, indicating that the next 12 to 18 months represent a critical window for optical communication adoption within server racks. Additionally, AWS, in partnership with QCOM, is providing financial support for 1.6T optical interconnects, further reinforcing this trend.

At the technology forum hosted by Deutsche Bank last week, Lumentum’s CEO Michael Hurlston also mentioned viewing Scale-In as the next major opportunity: this is a direction to watch over the next 12 to 18 months.

AVGO +3.0%

CEO Hawk Tan defended the $115B/$230B AI revenue path at the Goldman Sachs conference, stating that the constraint lies in power-ready sites rather than silicon, and that Broadcom has already secured capacity for 2027.

The key information lies in its value framework: for top-tier frontier models, each gigawatt of electricity generates approximately $30 billion in annual recurring revenue, with operating costs around $10 billion; token-based estimates show global inference costs at about $200 billion annually, while model revenue stands at $150 billion—frontier models, with half the token supply, capture at least 75% of the revenue, whereas open-source weight players spend roughly $100 billion to earn only $30 billion, indicating that value flows toward where intelligence continuously improves, leaving the second-place player with nothing.

He also refuted the cyclical criticism regarding the Apollo/Blackstone platform (claiming it creates demand through financing rather than being cyclical), and outlined a chip roadmap progressing from dual-chip to quad-chip to octa-chip.

Applied Materials (AMAT) rose 4.0%

Its CFO, Bryce Hill, stated at the Citi conference that all indicators still point to growth: the rolling eight-quarter forecast for top customers has increased each quarter this year; the wafer fab tracking system now covers over 100 fabs; capacity utilization for advanced logic and DRAM is nearing 100%; the bottleneck in capacity ramp-up lies in cleanroom construction, not demand.

The new information point is the estimation of new DRAM capacity additions, with annual wafer start volumes of 300,000 to 400,000 wafers per month over the coming years. Each 100,000-wafer start volume corresponds to approximately $10 billion in equipment investment, while upgrades require only about one-quarter of that investment—this is why he refers to it as a larger-scale investment cycle. He also noted that the downgrade in HBM specifications has a neutral impact on equipment demand, as the number of systems paired with each memory chip is increasing rather than decreasing.

ALAB -7%

Negative spillover from the QCOM/AWS deal: Amazon is its largest customer, and the QCOM deal targets internal network bottlenecks using jointly developed optical interconnect technology with speeds up to 1.6T, leveraging Qualcomm’s SerDes and optical DSP technologies—exactly within ALAB’s signal conditioning/interconnect domain (Aries, Taurus, Scorpio). Amazon already holds warrants for ALAB, contingent on procurement reaching $65 billion, but QCOM has signed a parallel $60 billion warrant, causing some investors to worry this may lead to market share erosion.

SNDK -0.1%

Chief Financial Officer Luis Visoso told Citigroup that the new business model has covered 50% of storage capacity this fiscal year and is expected to cover two-thirds by fiscal year 2028, with a gross margin floor of approximately 80%. Additionally, two contracts have been reopened as customers seek to increase their procurement volumes and extend contract durations. Supply growth is entirely driven by process node advancements, with no additional wafer capacity added, and industry storage capacity growth remains at a mid-to-high level of 15%-17%.

His view is that, by definition, demand cannot exceed supply. The incremental insight is that the KV cache demand model behind high-bandwidth flash has been raised in each recent revision, with related products scheduled for delivery to customers in 2027; $4.5 billion of the $5 billion in free cash flow from the last quarter was used for stock buybacks.

ANET increased by 0.6%

Chief Technology Officer Ken Duda and Chief Financial Officer Chantelle Breithaupt analyzed the rationale behind the raised guidance at the Goldman Sachs conference: supply chain improvements combined with order visibility across the two quarters through August, while the $9.7 billion in procurement commitments, nearly tripled, signal demand rather than inflation.

Horizontal scaling accounts for about 30% of this year’s $3.5 billion AI target, as customers cannot secure sufficient power at a single location; vertical scaling is an untapped market, with ANET holding nearly no share, expected to ramp up by end of 2027 and scale significantly by early 2028. Price increases apply only to memory-intensive bill of materials in future orders, aiming to maintain neutral margins. Regarding new cloud providers, the CFO’s principle is upfront payment or exit, as not all new cloud providers will succeed.