AI adoption constraints imposed by organisational speed, the need to rebuild software products around measurable value, the competitive position of Chinese open-weight models, and the practical friction of system validation formed the core of the fifth panel at the HKMA Investment Forum in Hong Kong on 4 November 2026. Moderated by Samson Wong, Chief Investment Officer, Private Markets, Exchange Fund Investment Office at the Hong Kong Monetary Authority, the discussion brought together Kenneth Hao of Silver Lake, Duane Kuang of Qiming Venture Partners, Deven J. Parekh of Insight Partners and David Siegel of Two Sigma.
David Siegel said expectations that large language models would transform businesses overnight had been unrealistic. The technology continues to evolve, but its deployment depends on people changing how work is organised and on processes being re-engineered. Those adjustments move at the speed of a human organisation, not the speed of a computer. Siegel argued that the volume of experimentation already under way indicates relatively rapid adoption, yet tangible bottom-line results take time to appear. He characterised the common framing of the question as unfair.
Kenneth Hao offered a different emphasis. In the boards and companies with which he is involved, AI is the first and last topic of discussion. Management teams must be fully committed before clean return-on-investment statistics become available. Measurement remains difficult, especially where AI drives both revenue and cost transformation. CIOs are already confronting steep token bills and beginning to differentiate use: frontier models deliver substantial productivity gains for the top 20 per cent of developers, while the remainder can be served by less expensive alternatives. He expects further rationalisation of token economics over the course of this year.
Deven J. Parekh said enterprise software companies must treat AI as a core product issue rather than an incremental feature. Many larger private-equity-backed software businesses are still led by executives with sales backgrounds in what has become a product-oriented market. Parekh described instructing one portfolio company that an entire subsequent board meeting would be devoted solely to AI and that a working product demonstration, not a strategy paper, was required.
He said the same company moved from zero AI-related revenue two years earlier to figures he stated as fifty million dollars by the end of this year and one hundred million dollars the following year, from a product that did not exist 18 months previously. Successful deployment, he argued, sometimes requires accepting that the existing product architecture is no longer relevant. Parekh used the phrase “burn the boat” to describe the need to rebuild around customer pain points. Those pain points now include labour cost as well as technology spend. In one portfolio example in customer support, the company was taking over a function rather than simply selling software, allowing the customer to compare exact costs and producing a rapid return on investment.
Duane Kuang said Chinese models are close in capability to leading United States models. Open-weight releases serve two purposes: they function as a marketing tool that allows independent testing and validation, and they build trust because an open model is easier to verify than a closed one, particularly in sensitive domains such as cybersecurity.
I am always surprised that people accepted this brand new technology, large language models, to transform business overnight. That was a complete unrealistic expectation.
David Siegel, Two Sigma
Once the decision to release open weights is taken, the business models resemble those of leading United States laboratories. Companies generate revenue through API and token services, where performance and price competitiveness matter, and through vertical applications built on their own models. Kuang observed that Chinese groups tend to place relatively greater emphasis on the end application than on pure model access. The valuation gap relative to United States peers, he said, reflects the current state of revenue and profitability rather than the open-weight strategy itself.
Responding to a question that referred to recent incidents, including an OpenAI disclosure that agents had escaped a testing environment, Siegel characterised such events primarily as cybersecurity problems. He said the industry would need to improve its cybersecurity models.
A larger practical issue, he argued, arises when agentic systems are deployed inside companies. Executives responsible for deployment must be able to validate that the system will perform as expected and understand where problems may arise. In a traditional organisation, accountability runs through human managers who can be questioned or replaced. Agentic systems require new validation tools and adjustments to how a business is run. Siegel said these problems would generate further headlines and would take more than six months or a year to resolve, creating friction that slows deployment without ultimately preventing transformation.
Kenneth Hao discussed Silver Lake’s acquisition of a 51 per cent stake in Altera, the programmable-chip business previously owned by Intel. The thesis rested on franchise quality and timing. Programmable logic sits alongside CPUs, GPUs and the custom inference chips now being designed by hyperscalers and frontier model laboratories. Hao noted that the accelerator market is already fragmenting and that start-up competition in pure GPUs is likely to face rising barriers unless capacity shortages persist.
Duane Kuang added that significant activity in China is occurring at the inference and edge layers, where technology requirements are less extreme. Alternative memory architectures that reduce reliance on high-bandwidth memory are also attracting capital.
The absence of near-term return-on-investment disclosure does not automatically indicate wasted spend; organisational absorption rates differ from technology release cycles, as Siegel emphasised. Second, software and enterprise technology holdings require product-level scrutiny: companies still treating AI as an incremental feature rather than a potential replacement of core architecture face higher disruption risk. Third, China exposure in AI requires separate assessment of model capability, commercial traction and the layers of the semiconductor stack where domestic suppliers can compete. Fourth, validation frameworks for agentic systems will influence the pace at which large institutions themselves deploy these tools.