If you want to provision data center sites with the necessary power by 2028, you have to start construction work and obtain permits now.
Article index:
- Historic financial forecasts at Goldman Sachs conference
- Infrastructure crisis and physical power limits
- Jalapeno chip with OpenAI and major partnerships
- Decade-long extended partnership with Google
- Economic value concentration and closed model profits
- Massive cash flows and buyback decisions
- FAQs
Historic financial forecasts at Goldman Sachs conference
Hock Tan, CEO of global company Broadcom, presented the most detailed and precise vision for his company’s growth trajectory in the artificial intelligence sector during his participation in the Goldman Sachs Technology and Communications Conference on Tuesday. Tan predicted that his company’s artificial intelligence revenues would record about $115 billion in fiscal year 2027, rising to $230 billion in fiscal year 2028, in figures he described as “exceptional by all standards” but constrained by physical supply chains rather than increasing demand levels.
These forecasts, whose broad outlines were revealed in the third-quarter 2026 earnings report on September 2, reflect a full doubling of artificial intelligence revenues for each of the next two fiscal years, up from about $58 billion projected for the full year 2026.
Infrastructure crisis and physical power limits
Tan explained to attendees that the strict constraint holding back the rapid pace of growth in the artificial intelligence revolution is not related to a shortage of semiconductor supply or slow manufacturing, but goes directly back to physical infrastructure: represented by the scarcity of data center sites ready to be supplied with sufficient electricity, the complexities of obtaining environmental permits, and the lengthy time periods required for engineering construction.
Tan warned: “If you want to provision a massive data center with the necessary electric power by 2028, you actually have to start construction and building work inside those sites right now without any delay,” emphasizing that providing electrical grids and generation stations has become the industry’s biggest challenge.
Jalapeno chip with OpenAI and major partnerships
Broadcom’s activity in this sector is based on the design and manufacturing of custom computing accelerators known as XPUs, which are built in close engineering cooperation with six of the world’s leading companies developing cutting-edge models. Tan highlighted joint cooperation with OpenAI, noting that the two parties successfully developed a custom artificial intelligence chip known internally as Jalapeno in approximately one year.
The Broadcom chief confirmed that the Jalapeno chip “operates with performance that matches, if not exceeds, the best and latest general computing accelerators available globally,” revealing that his company is already working on developing two subsequent generations of this advanced chip to support future model training.
Decade-long extended partnership with Google
Tan touched upon the long-standing partnership uniting Broadcom with Google, a historic relationship spanning over a full decade that began since the design of the first version of Tensor Processing Units (TPUs). The two parties crowned this partnership last April by signing a long-term agreement covering the supply of processing units and advanced data networking products through 2031.
This alliance guarantees securing ultra-efficient silicon flows for Google data centers around the world, enabling it to compete with cloud computing giants while maintaining energy consumption efficiency.
Economic value concentration and closed model profits
Tan put forward a striking analytical vision regarding the distribution of economic value, considering that profits and margin returns in the artificial intelligence sector will concentrate among the most efficient closed model developers rather than being distributed evenly across supply chains. He explained via an illustrative example that providing computing capacity equivalent to one gigawatt of power can generate about $30 billion in recurring annual revenues against operating costs of only $10 billion, leaving the lion’s share of profits to model developers and their commercial applications.
Tan provided a comprehensive overview of the market, revealing that global token generation infrastructure costs about $200 billion annually against model-related revenues of about $150 billion, pointing out that leading closed models capture at least 75 percent of those revenues despite generating only half of the tokens, which raises question marks about the economic viability of open-weight models.
Massive cash flows and buyback decisions
Broadcom expects to end its fiscal year 2026 with an unprecedented historic cash balance, and achieve free cash flows exceeding the middle of the $40 billion range in 2027. Tan explained that the company’s debt of $56 billion was issued during periods when interest rates were low, making early repayment economically unfeasible.
He noted that the board of directors will discuss capital allocation strategies at its scheduled meeting during the month of December, where increasing cash distributions and share repurchase programs are the most likely options, concluding by saying: “We will generate massive cash liquidity, and that will magnify even further in 2028.”
FAQs
Question: What are Broadcom’s artificial intelligence revenue forecasts for 2027 and 2028?
Answer: The company expects to achieve $115 billion in 2027 and double it to reach $230 billion in 2028.
Question: Why is power availability considered the biggest obstacle to the growth of intelligence data centers?
Answer: Because building generation stations and provisioning sites with electricity requires multi-year construction periods and long environmental permits.
Question: What is the Jalapeno chip developed by Broadcom with OpenAI?
Answer: It is a high-efficiency custom computing accelerator developed in one year that rivals the best global chips available in the market.
