AVGO Lowers 2027 AI Expectations but Maintains Buy; NVDA and Terafab Accelerate Vertical Integration of the AI Supply Chain

Overview

The report centers on Broadcom's earnings preview and extends its analysis to NVIDIA's vertical integration into AI infrastructure and model development, Broadcom's participation in AI computing-capacity financing, and SpaceX Terafab's recruitment and construction progress. UBS lowered its F2027 AI revenue forecast and target price for Broadcom but maintained its Buy rating, believing that the timing of long-term AI demand realization depends more on whether powered land and building shells are ready than on the availability of compute chips themselves.

Core views

Regarding Broadcom's results, UBS expects modest upside to FQ3 revenue for the quarter ended July and expects the company to guide to approximately $34B in FQ4 revenue for the quarter ended October. FQ3 AI revenue is expected to be $16B, rising to $21B-$22B in FQ4; custom compute is expected to continue growing faster than the AI networking business, primarily driven by expanding Google TPU and Anthropic deployments. Key topics listed for the earnings discussion include the custom ASIC revenue trajectory, the ramp of TPU and other XPU programs, Anthropic TPU shipments, progress on OpenAI Jalapeño in FY26 and beyond, Broadcom's position in Google's TPU program, AI networking and CPO opportunities, and XPV financing progress. Anthropic is viewed as an important source of near-term incremental growth. UBS believes FQ3 is the first quarter in which its TPU shipments become meaningful, estimating a contribution of approximately $2.5B for the quarter, rising to approximately $3.3B in FQ4 and approximately $9.3B in cumulative shipment revenue in C2026. Google Cloud-related AI revenue is expected to increase by approximately $2B quarter over quarter in FQ3 and by another approximately $1.5B quarter over quarter in FQ4. Although Google is advancing its in-house chip collaboration with MediaTek and MRVL is participating in a non-TPU AI inference accelerator project, UBS still expects Broadcom to emphasize that it remains the "plan of record" for Google's TPU program and believes the MRVL project will not erode Broadcom's share or opportunity in that program. The pace among other custom compute customers is uneven. UBS expects META custom compute shipments to remain broadly flat through F2026, with more pronounced growth emerging only in F2027; OpenAI is expected to contribute approximately $1.25B and $12B of revenue in C2026 and C2027, respectively, while AAPL's ramp is comparatively slower. Because channel checks indicate that previous TPU shipment expectations were overly aggressive, UBS reduced its Broadcom F2027 AI revenue forecast by approximately $5B to approximately $130B and revised its CY27 TPU shipment forecast to 6.4MM units. The report expects the company not to update its previous framework of "more than $100B in F2027 AI revenue," instead steering investors more toward F2028, when AI revenue could increase significantly to approximately $200B. The report's TPU forecast table presents figures from C2024 through C2028: GCP TPU revenue of $7,676, $13,204, $30,838, $58,538, and $73,238, respectively; Anthropic revenue of —, $1, $9,300, $18,500, and $14,000, respectively; total TPU revenue of $7,676, $13,204, $40,138, $77,038, and $87,238, respectively; and UBS total AI revenue forecasts of $12,982, $24,428, $70,702, $152,364, and $209,619, respectively. This calendar-year table uses a different period basis from the approximately $130B F2027 forecast in the main text, and the report retains both sets of estimates. UBS believes these results are unlikely to end the market debate surrounding Google and custom chips, and that the actual timing of the AI revenue ramp is constrained more by the availability of powered land and building infrastructure than by compute-chip supply. Regarding NVIDIA, the report believes its business model is evolving from a GPU supplier toward deeper participation in land, power, and building shells, or the LPS layer. Most AI capacity is currently procured by hyperscale cloud providers or specialized cloud service providers, but AI labs increasingly want to control long-term infrastructure. Under the PORTS-Pike arrangement, OpenAI remains the tenant and operator, while NVIDIA is jointly investing $1.5B with SB Energy and providing up to $105B in residual-value or credit support to help secure the underlying land, power, and buildings. Initial planned capacity is 4.25GW, potentially expandable to approximately 8GW. NVIDIA estimates that each generation of AI factory systems at the campus could deploy approximately 1.5MM GPUs, corresponding to approximately $150B-$200B in revenue, with multiple upgrade cycles also possible over the campus's lifetime. UBS views NVIDIA's investments in SB Energy, Lancium, and Cloverleaf Infrastructure as part of establishing GPU "capacity corridors" and believes it could eventually become a developer, or even an operator, of AI factory assets. Land, power, and buildings are becoming strategic inputs in the AI supply chain, and NVIDIA's scale and financial resources can both help secure these inputs and enhance its visibility into future AI demand. NVIDIA's transaction with Poolside further demonstrates its expansion into model-development infrastructure. NVIDIA agreed to pay $6B for a non-exclusive license to the Poolside Model Factory platform and hire 109 employees; it also invested $1B in the remaining company at a $12B pre-money valuation. UBS believes the transaction's focus is not on purchasing a particular model, but on acquiring the "machine" used to build models, including software, training systems, and engineering processes. The platform is used to develop the Lagunas open-weight MoE model for agentic tasks such as coding and is expected to accelerate NVIDIA's progress in open-source foundation models and bring in key talent without requiring a full acquisition of Poolside. If open models are both trained on NVIDIA hardware and optimized for NVIDIA's inference platform, this could create a positive feedback loop for AI demand; however, the report specifically notes that the $6B license fee is substantial, implying that NVIDIA assigns considerable standalone value to the model-development platform and personnel. Broadcom may also arrange up to approximately $100B in AI infrastructure financing for Anthropic and other customers, including $60B-$70B in senior secured financing partially guaranteed by Broadcom and approximately $30B in subordinated debt, with Blackstone and Apollo also expected to participate. UBS estimates that this headline amount corresponds to approximately 2GW and may be related to the AI XPV framework launched in June 2026; the framework was established by Broadcom, Apollo, and Blackstone with the goal of supporting more than 20GW of AI compute deployments by 2028, beginning with financing for Anthropic's first project exceeding 1GW. UBS believes the new financing further validates Broadcom's ability to help customers raise capital for large-scale XPU deployments, representing an incremental positive, but most buy-side models may already reflect this factor, and approximately 2GW remains relatively small compared with the XPV framework's overall opportunity exceeding 20GW. SpaceX Terafab indicates that vertical integration of the semiconductor supply chain is entering the practical construction-preparation stage. UBS Evidence Lab tracks job openings and construction activity at the Grimes County, Texas campus: Terafab-related openings reached 44 in July, mostly targeting senior engineers with high-volume manufacturing or fab-operations experience and covering major processes such as lithography, deposition, dry etching, and epitaxy. This suggests it may be building a relatively comprehensive engineering organization rather than focusing on only a few process modules. Some positions appeared concurrently in Austin, Texas and Hsinchu, Taiwan, China; UBS speculates that SpaceX may hope to attract TSMC talent, although how recruitment in Taiwan will directly support the Texas fab's ramp over the coming years remains unclear. Recruitment already includes memory design and memory integration engineers, indicating that Terafab is willing to initiate in-house memory manufacturing, but the source of memory IP remains a key unresolved issue, and existing memory suppliers may have limited incentive to license IP to it. Following the project's formal announcement on August 6, cleanroom-construction-related openings increased significantly and became the largest recruitment category; however, satellite imagery from August 2026 shows only limited site clearing and tree removal, indicating that physical infrastructure remains at a very early stage. Regarding Broadcom's valuation, UBS modestly lowered its target price from $485 to $470 due to the reduction in its F2027 AI revenue forecast but maintained its Buy rating. The valuation uses a business-segment EV/FCF approach: approximately 12x EV/FCF is applied to the CY27E Software free-cash-flow contribution of $35.1B, while approximately 30x EV/FCF is applied to the CY27E Semis free-cash-flow contribution of $65.6B, with the multiple assumptions unchanged. The report's valuation chart also shows Broadcom trading at approximately 21x P/E and approximately 20x EV/FCF, representing an approximately 1x P/E premium to the S&P 500 but an approximately 14x P/E discount to the semiconductor sector.

Analysis framework

UBS first uses customer and supply-chain checks to establish Broadcom's AI revenue and TPU shipment forecasts by customer and project, then connects the near-term quarterly trajectory with the longer-term F2027-F2028 ramp and adjusts its target price accordingly. For NVIDIA, the report breaks down the land, power, and building financing arrangements and the structure of the Poolside transaction to assess the direction of its vertical integration; for Terafab, it combines data from Tesla and SpaceX recruitment websites, job-skill classifications, and satellite imagery to monitor construction progress. Finally, the report uses business-segment EV/FCF multiples to value Broadcom and cross-compares P/E, EV/FCF, and multiples relative to the broader market and semiconductor sector.

Methodology notes

• Valuation MethodSOTP Segment Valuation

  Broadcom Business-Segment Valuation

  The report separately estimates CY27E free cash flow for the Software and Semis businesses and applies different EV/FCF multiples to the two segments, reflecting differences in their growth and business characteristics to derive the overall target price.

• Valuation Method

  EV/FCF Valuation

  EV/FCF compares enterprise value with free cash flow; this report applies approximately 12x to Broadcom Software and approximately 30x to Semis, deriving a $470 target price.

• Valuation MethodPE/PEG valuation

  P/E Multiple Valuation and Relative Valuation

  The report states that NVIDIA's target price uses a P/E multiple and compares Broadcom's approximately 21x P/E with its relative premium or discount to the S&P 500 and semiconductor sector.

• Industry/Sector Analysis FrameworkSupply-demand framework

  Infrastructure Constraints on AI Computing-Capacity Deployment

  The report views land, power, and building shells as critical supply inputs for expanding AI capacity and concludes that Broadcom's AI revenue ramp is constrained more by the pace at which powered facilities become ready than by chip availability.

• Industry/Sector Analysis Framework

  Recruitment Data and Satellite Imagery Monitoring

  UBS Evidence Lab directly collects data from Tesla and SpaceX recruitment websites, cleans anomalies and delisted positions, and combines the job structure with satellite imagery to assess Terafab's technology roadmap, recruitment intensity, and construction stage.

Asset mapping & comparison

Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).

• Broadcom Inc. (AVGO)

  Google, Anthropic, and other custom XPU projects are driving AI revenue growth, while the company also helps customers build AI infrastructure through financing structures such as XPV.

  Strengths

  Custom compute is growing faster than the AI networking business; UBS still views the company as the established partner of record for Google's TPU program; it has the ability to help customers secure funding for large-scale deployments.

  Weaknesses

  UBS lowered its F2027 AI revenue and CY27 TPU shipment forecasts, projects such as META and AAPL are ramping more slowly, and revenue realization is also constrained by the construction of powered facilities.

  Comparison

  Its approximately 21x P/E is approximately 1x above the S&P 500 but approximately 14x below the semiconductor sector; custom compute is expected to continue growing faster than the AI networking business.

  Risks

  M&A could affect free cash flow and the payout ratio; market tightening could reduce dividend-based valuation; the company also faces market-share, macroeconomic, geopolitical, foreign-exchange, and tax risks.

• NVIDIA Corp. (NVDA)

  Through projects such as PORTS-Pike, the company is securing land, power, and buildings, while the Poolside transaction provides model-development infrastructure and talent, enabling vertical integration into the AI factory value chain.

  Strengths

  Its scale and financial resources help secure scarce infrastructure; a positive demand feedback loop may form among GPUs, open models, and inference platforms; visibility into future AI demand could improve.

  Weaknesses

  Poolside's $6B license fee is substantial, while the infrastructure arrangements also create exposure of up to $105B in residual-value or credit support.

  Comparison

  The report provides no direct peer valuation comparison but believes NVIDIA has unique advantages in scale and financial resources for securing land, power, and buildings.

  Risks

  It faces competition from AMD in GPUs, ARM application-processor competitors, and Intel's MIC architecture, and is affected by semiconductor industry cycles and corporate earnings trends.

• SpaceX Terafab

  The project aims to increase vertical integration within SpaceX's silicon supply chain and could generate long-term demand for the broader semiconductor equipment supply chain.

  Strengths

  Recruitment covers major processes including lithography, deposition, dry etching, epitaxy, and memory design and integration, indicating plans to build relatively comprehensive engineering capabilities.

  Weaknesses

  Satellite imagery indicates that physical construction remains at the limited site-clearing stage, it is unclear how recruitment in Taiwan, China will support the Texas facility, and the source of memory IP remains unresolved.

  Risks

  Existing memory suppliers may lack incentives to license IP, while the speed at which infrastructure and talent become available could affect the actual fab construction and mass-production timeline.

Key data

• Broadcom Earnings Release DateAfter market close on September 2AVGO earnings release timing stated in the report
• Broadcom FQ4 Revenue Guidance ForecastApproximately $34BUBS's estimate of the company's revenue guidance for the quarter ended October
• Broadcom AI RevenueFQ3 $16B; FQ4 $21-22BExpected to grow as TPU and Anthropic deployments expand
• Anthropic TPU RevenueFQ3 approximately $2.5B; FQ4 approximately $3.3B; C2026 approximately $9.3BFQ3 is viewed as the first quarter of meaningful shipments
• Incremental GCP-Related AI RevenueFQ3 quarter over quarter approximately +$2B; FQ4 quarter over quarter approximately +$1.5BGoogle deployment remains strong
• OpenAI Revenue Contribution ForecastC2026 approximately $1.25B; C2027 approximately $12BBroadcom custom compute customer forecast
• Broadcom F2027 AI Revenue ForecastApproximately $130BLowered by approximately $5B
• CY27 TPU Shipment Forecast6.4MMAdjusted because the previous forecast was overly aggressive
• Broadcom F2028 AI Revenue OutlookApproximately $200BExpected to increase significantly from F2027
• PORTS-Pike Initial and Potential Capacity4.25GW, expandable to approximately 8GWOpenAI is the tenant and operator
• NVIDIA Infrastructure Support$1.5B investment and up to $105B in residual-value/credit supportTo help secure land, power, and building infrastructure
• PORTS-Pike System Scale per GenerationApproximately 1.5MM GPUs, corresponding to approximately $150-200B in revenueNVIDIA estimates that there could be multiple upgrade cycles over the campus's lifetime
• Poolside Transaction$6B license fee, 109 employees, and an additional $1B investmentThe remaining company has a pre-money valuation of $12B
• Broadcom Potential Financing PackageUp to approximately $100BIncluding $60-70B in senior secured financing and approximately $30B in subordinated debt, corresponding to approximately 2GW
• XPV Framework TargetMore than 20GW by 2028Initially supporting Anthropic's first deployment exceeding 1GW
• Number of Terafab Job Openings44July 2026, primarily recruiting senior manufacturing and fab-operations engineers
• Broadcom Target Price$470Lowered from $485, with the Buy rating unchanged
• Broadcom Business-Segment ValuationSoftware approximately 12x EV/FCF; Semis approximately 30x EV/FCFCorresponding CY27E free-cash-flow contributions are $35.1B and $65.6B, respectively
• Broadcom Market ValuationApproximately 21x P/E; approximately 20x EV/FCFP/E is approximately 1x above the S&P 500 and approximately 14x below the semiconductor sector

Impact & implications

The report believes Broadcom's near-term AI revenue remains driven by Google and Anthropic deployments, but the pace of F2027 TPU shipment and revenue realization needs to be recalibrated, with the actual bottleneck shifting toward power, land, and buildings. Through financing, credit support, and investments in model-development platforms, NVIDIA is expanding both upstream and downstream in the AI value chain, potentially strengthening its control over future GPU demand and deployment capacity. Terafab's broad process-related recruitment has long-term significance for the semiconductor equipment supply chain, but construction remains at an early stage, particularly as the source of memory IP and the ramp pathway for the Texas facility remain unclear.

Risks

• Broadcom's continued M&A activity could change free cash flow and the payout ratio, thereby affecting the valuation basis for its target price.
• Tighter market conditions could affect Broadcom's dividend yield and cause valuation downside.
• Semiconductor and equipment companies face risks from macroeconomic conditions, geopolitics, international trade, foreign exchange, taxation, and changes in market share.
• New technologies or business models could alter industry unit volumes, average selling prices, and revenue trajectories.
• NVIDIA faces competition from AMD in GPUs and professional visualization, ARM application-processor competitors, and Intel's MIC architecture.
• The source of Terafab's memory IP remains unclear, and existing memory suppliers may lack incentives to license it.

What to watch

• Watch Broadcom's commentary on the custom ASIC revenue trajectory and the ramp of TPU and other XPU programs.
• Watch Anthropic TPU shipments and progress on OpenAI Jalapeño in FY26 and beyond.
• Watch how Broadcom responds to the implications of MediaTek's work on Google's in-house chips and MRVL's new accelerator project for its TPU position.
• Watch the scale and pace of development in AI networking, scale-up, and CPO opportunities.
• Watch XPV financing progress and whether the approximately $100B potential financing package can translate into actual deployment capacity.
• Watch whether powered land and building shells become ready on schedule, as this is a key constraint on the AI revenue ramp.
• Watch Terafab's cleanroom-construction recruitment, construction progress in satellite imagery, the source of memory IP, and the actual ramp of the Texas site.
• Watch NVIDIA's vertical-integration progress from AI factory asset development toward operations.