Tong's Portfolio - Open

This article first appeared in The Edge Malaysia Weekly on July 27, 2026 - August 2, 2026

Running frontier artificial intelligence (AI) is not cheap. For a time, “token maxxing” was the corporate mantra du jour. Earlier this year, companies encouraged employees to maximise their AI usage by consuming as many tokens as possible. Tokens, in this context, refer to the basic unit of computing. The assumption was that greater AI utilisation would translate into greater productivity — never mind that the relationship between AI usage and output quality is not necessarily linear.

Since then, AI costs have skyrocketed. Newer and more complex reasoning models, alongside an expanding suite of AI tools, have driven up token consumption per query. The result has been soaring enterprise AI bills. As Axios covered in late May, one company reportedly accumulated a US$500 million bill from Anthropic’s Claude in a single month alone.

In response, a number of companies including Uber, Cisco and Walmart have begun introducing internal guard rails to prevent runaway inference spending. Even Big Tech has not been immune to AI sticker shock. In mid-May, Microsoft announced the cancellation of its internal Claude Code licences, in part to rein in operating costs ahead of the company’s new financial year.

This newfound parsimony undoubtedly poses a problem for frontier labs, which have staked much of their commercial ambitions on enterprise adoption. A bigger problem yet is the fact that most businesses do in fact have a cheaper alternative to consider: open-weight models.

Readers following AI’s latest developments will likely have come across the terms “open source” and “open weight”, often used interchangeably when discussing alternatives to proprietary models such as ChatGPT and Claude. There is in fact a difference between the two.

The confusion stems from the origins of the term “open source” — a concept that predates modern AI by several decades. In the 1980s, software developers advocating for greater transparency amid the growing dominance of proprietary software began making their source code publicly available for anyone to inspect, modify and redistribute. The term “open source” was later adopted to describe this collaborative approach to software development.

AI models, however, add another layer of complexity. Unlike conventional software where openness centres on the source code, AI models comprise of three broad components: the source code which defines the model’s underlying architecture; the training data used to train it; and the model weights that encode what the model has learnt.

Accordingly, truly open-source AI models, such as Ai2’s OLMo, offer the highest degree of transparency by making all three components available to the public. In practice, however, most AI labs that contribute to the open ecosystem release only the model weights. Examples including DeepSeek’s models, Meta’s Llama family, and certain versions of Alibaba’s Qwen family are thus more accurately described as open-weight models (see Table 1 for summary).

This distinction is more than just a matter of semantics, given that open-weight models are sometimes distributed under bespoke licences that may impose additional stipulations such as revenue-sharing conditions. MiniMax M models, for instance, use a modified MIT Community License that requires written authorisation and revenue sharing for some commercial deployments. In practical terms, this illustrates one avenue through which AI labs can monetise open-weight models while still making their weights publicly available.

Nevertheless, both open-source and open-weight models share an important practical characteristic: beyond conventional chatbot or application programming interface (API) access, these models can also be downloaded and deployed independently. By contrast, closed models are accessible only through provider-hosted chatbots or APIs. This distinction carries significant economic implications, which we explore further below.

When it comes to open-weight models, DeepSeek’s R1 release last January is widely heralded as AI’s “Sputnik moment” for delivering near-frontier level capabilities at a fraction of the costs of its closed peers. Subsequent releases have only continued to push the frontier of open-weight models. Z.ai’s GLM-5.2 and Moonshot AI’s Kimi K3, released in June and July respectively, are the latest in a string of increasingly capable open-weight models. Kimi K3, in particular, dazzled the tech world with benchmark performances, placing it almost on par with leading proprietary models (see Chart 1).

As with DeepSeek R1, the emergence of these low-cost, computationally efficient models has once again reignited questions over whether Silicon Valley’s AI spending spree is justified.

This is certainly a fair question. While proprietary models continue to hold a narrow lead on frontier benchmarks, real-world adoption rarely hinges on capability alone. Cost, customisability and deployment flexibility often matter just as much. More importantly, frontier-level intelligence is simply unnecessary for most enterprise workloads. AI that is “good enough” is, in practice, often good enough.

This is precisely the value proposition that Chinese AI labs have embraced. Since the second half of 2025, the open ecosystem has become primarily defined by Chinese models (see Chart 2). Rather than competing solely on benchmark supremacy, Chinese AI labs have increasingly focused on open-weight releases, encouraging developers to adapt, fine-tune and build upon their models.

This strategy may yet prove decisive in the AI race, where leadership is likely to depend not only on who builds the most capable models, but on who succeeds in driving their widespread adoption. The rise of open-weight models underscores this exact point: technological superiority alone does not guarantee market leadership. What matters just as much is the ability to translate innovation into widespread, low-cost deployment.

Indeed, the current scale of China’s open-weight ecosystem suggests the effectiveness of this approach. Today, Alibaba’s Qwen family boasts the world’s largest open-weight ecosystem. In February 2026 alone, Qwen generated more than 150 million downloads, more than double the next eight open-weight contenders combined (see Chart 3).

For investors, a separate question is equally worth examining: As open-weight models continue to gain traction, where exactly will the money go?

As it stands, the economics of proprietary models are relatively straightforward. Companies such as OpenAI and Anthropic retain control over their model weights and typically provide access through managed services, either directly or via select cloud partners. Model and compute are bundled into a single service, with customers paying through subscription or usage-based pricing.

When it comes to open-weight models, however, these two layers are separated. Once downloaded (for free), open-weight models can be deployed almost anywhere: through a third-party inference platform; within an enterprise’s own data centre; or in the case of smaller models, even on a sufficiently powerful laptop. The user decides where the model runs.

This seemingly subtle distinction carries profound commercial implications. Because the underlying model can be freely accessed, open-weight AI commodifies the model layer. At the same time, because the AI model can be deployed across multiple providers and environments, customers gain greater flexibility to move workloads across to whichever platform offers the best value proposition. This competitive pressure drives down inference costs for open-weight models (Chart 4 illustrates this point clearly).

Consequently, as open-weight AI moves further into the mainstream, the model layer becomes less differentiated, with value shifting towards the broader ecosystem that enables its deployment. As discussed below, the following sectors illustrate where value may increasingly accrue, though they are by no means exhaustive.

Among the clearest beneficiaries of the open-weight shift are neoclouds. Prominent names within this sector include CoreWeave, Nebius and Iren. Unlike traditional hyperscalers, neoclouds focus almost exclusively on high-performance computing. This narrower focus allows them to optimise specifically for AI demand — enabling faster deployment and materially lower costs. Research from Uptime Institute, for instance, found that renting Nvidia H100 GPUs through neocloud providers amounted to cost savings of up to 66%. For cost-conscious businesses, the attraction is obvious.

That said, this does not imply that hyperscalers will become irrelevant in the AI compute landscape. For larger enterprises, AI infrastructure constitutes only one component of their broader technology stack. Hyperscalers therefore remain attractive to companies seeking an integrated platform that combines compute infrastructure, enterprise applications, developer tools and other cloud services under a single provider.

Furthermore, the neocloud operating model is not without risk either. The sector’s poster child CoreWeave illustrates both the opportunity and fragility inherent in this business. While revenue more than doubled in the latest quarter, net losses widened even more sharply, up 433% year on year amid an aggressive debt-funded expansion. The company’s debt-to-equity ratio stands at a whopping 475% as of its latest reporting period. At the same time, competition is intensifying with the entry of new competitors. Concerningly, Meta — CoreWeave’s largest customer — is among them, having recently announced plans to enter the AI compute market.

Beyond cost savings, open-weight models also significantly lower the barrier to enterprise adoption in two important ways. First, by allowing companies to adapt existing models instead of building from scratch, they reduce the upfront investment required for businesses to develop AI capabilities.

Second, because open-weight models can be deployed on-premises or in hybrid environments, companies are able to retain greater control over where their data is stored and processed. For industries where data cannot leave the organisation’s infrastructure — whether due to regulation, security concerns or confidentiality — this flexibility can be a decisive advantage.

The challenge, of course, lies in implementation. Managing a self-hosted AI model requires considerably greater technical expertise. This is where enterprise IT vendors such as Dell, Hewlett Packard Enterprise (HPE) and Penguin Solutions stand to benefit. The added complexity of private deployments expands the addressable market for vendors that can abstract away the operational burden for their customers. HPE’s GreenLake platform, for instance, offers enterprises a turnkey solution for managing private clouds, handling much of the complexity involved in deployment, management and ongoing maintenance.

On a separate note, the struggle to access, integrate and extract value from enterprise data continues to be the leading reason behind the underperformance of corporate AI initiatives (see Chart 5).

This challenge could become even more pronounced as AI models become cheaper and more accessible. After all, the ability to deploy AI will matter little if enterprises cannot provide these systems with relevant, high-quality input.

This creates an increasingly important role for companies such as Snowflake, MongoDB and Databricks, which sit within the enterprise data stack. By helping organisations organise, govern and make sense of their data, these platforms provide the foundation needed for AI systems to access the relevant information and perform capabilities such as knowledge retrieval and workflow automation.

Data security represents another critical layer as well. As enterprises grant AI systems access to increasingly sensitive information, companies such as Rubrik and Cohesity ensure that organisational data remains resilient against threats ranging from accidental loss to ransomware attacks.

Finally, as deployment costs fall and implementation becomes more flexible, open-weight models will likely expand the range of workflows in which AI can be deployed. In this scenario, the companies that integrate AI most effectively will stand to capture the greatest gains through higher productivity and stronger margins.

The ultimate beneficiaries of cheaper intelligence may therefore not be those building the underlying infrastructure or enabling its deployment, but their customers. That is, the companies that successfully apply AI to reshape their own operations, now at a fraction of the previous cost.

After all, lest we be momentarily led astray by the metrics that have come to define the AI race, the true measure of progress will not be found in benchmark scores or compute demand. Rather, AI’s value will ultimately be determined by how effectively it is put to work to generate tangible real-world outcomes.

The Malaysian portfolio fell 0.6% for the week ended July 22. The two gaining stocks were Public Bank (+2.0%) and LPI Capital (0.4%), while the biggest losers were Hong Leong Industries (-3.2%), United Plantations (-1.9%) and Kim Loong Resources (-1.8%). Total portfolio returns now stand at 225.6% since inception. This portfolio is outperforming the benchmark FBM KLCI, which is down 6.5% over the same period, by a long, long way.

The Absolute Returns Portfolio also ended lower last week, down 1.4%. The loss pared total portfolio returns to 28.0% since inception. The top gainers include Berkshire Hathaway (+0.2%), Alibaba Group Holding (+0.1%) and Sun Hung Kai Properties (+0.1%) while the notable losers were Alphabet Inc - CL C (-7.6%), Talen Energy Corp (-5.7%) and Microsoft Corp (-1.3%). We added two stocks to the portfolio, US-based Thermo Fisher Scientific and Singapore Technologies Engineering, which is listed on the Singapore Exchange. The acquisitions reduced our cash holding to about 20% of total portfolio value.

The AI portfolio fell 1.0% over this same period. Total portfolio returns now stand at 23.3% since inception. The biggest gainers were Unusual Machines (+6.2%), Akamai Technologies Inc (+3.6%) and Naura Technology (+2.8%), while the top losers were Cadence Design Systems (-9.3%), Datadog (-7.1%) and Amazon.com Inc (-4.0%).

Disclaimer: This is a personal portfolio for information purposes only and does not constitute a recommendation or solicitation or expression of views to influence readers to buy/sell stocks. Our shareholders, directors and employees may have positions in or may be materially interested in any of the stocks. We may also have or have had dealings with or may provide or have provided content services to the companies mentioned in the reports.

Save by subscribing to us for your print and/or digital copy.

P/S: The Edge is also available on Apple's App Store and Android's Google Play.