We don’t often look to traditional finance for blockchain insights, but when CITIC Securities International raised Tencent’s capital expenditure forecast to HKD 215.7 billion for 2026 and HKD 260 billion for 2027, something clicked. This isn’t just a story about a Chinese tech giant doubling down on AI. It’s a signal about the concentration of computational power, and a mirror for the decentralized protocol world that’s trying to build an alternative. I’ve been tracking the intersection of AI and crypto since 2023, when I launched a prototype called TruthLayer on a Nairobi balcony. The numbers from Tencent’s Q2 report—core profit up 19%, AI capex surging, but depreciation costs eating into net profit—read like a case study in the tension between centralized scale and decentralized resilience. Let me unpack what this means for builders, investors, and anyone who believes that intelligence should be permissionless.
Context: The Tencent AI Machine and Its Financial Architecture CITIC Securities, a major Hong Kong-based brokerage, published its note on August 14, 2025, after Tencent’s second-quarter earnings slightly beat expectations. The headline: domestic gaming and advertising grew beyond forecasts, operating profit (excluding new AI product investments) rose 19% year-on-year. But the real story is the AI strategy. Tencent has now clarified four major AI pillars—likely covering large language models, AI-powered advertising, cloud AI services, and internal productivity tools. This clarity justifies the aggressive capex: HKD 215.7 billion for 2026, HKD 260 billion for 2027. Compare that to Tencent’s 2024 total revenue of about HKD 620 billion—over 40% of revenue will be plowed into infrastructure. The firm maintains a ‘Buy’ rating, though the target price dipped from HKD 632 to HKD 620, reflecting near-term depreciation pressure. Core net profit estimates for 2026-2028 are cut by 5-9%, with growth rates of just 2% and 3% in 2026 and 2027. The bear market sentiment in traditional finance is real: even a cash-rich giant like Tencent faces friction when it converts profits into compute.
From a blockchain perspective, this is fascinating. Tencent’s capex is essentially a bet on centralized compute—hundreds of thousands of GPUs, proprietary data centers, and closed-source models. The decentralized alternative—projects like Bittensor, Render Network, or Akash—currently command a fraction of that capital. According to CoinGecko, the total market cap of all AI-focused crypto tokens (including compute, data, and model tokens) is roughly $15 billion as of August 2025. That’s less than 7% of Tencent’s single-year capex. The asymmetry is staggering. But the bear market didn’t kill the thesis; it clarified the trade-offs. Centralized AI can scale faster, but it inherits single points of failure, censorship risks, and opaque governance. Decentralized AI is slower, but it offers verifiability, composability, and permissionless access. The question is whether the financial pressure from Tencent’s scale will suffocate the grassroots, or force it to evolve.
Core: The Technical and Values Analysis of AI Capex in a Decentralized World Let’s dig into the numbers. Tencent’s capex increase implies massive depreciation costs—servers, networking gear, and cooling systems have a 3-5 year lifespan. CITIC Securities explicitly lowered core net profit estimates because depreciation eats into reported earnings. This is a classic scaling dilemma: to win the AI race, you must spend now and accept a temporary dent in profitability. The same dynamic plays out in blockchain, but with a twist. When a Layer 2 project like Arbitrum or Optimism spends heavily on sequencer infrastructure or data availability, it’s often funded by token emissions or treasury reserves, not operating profits. The cost of capital is different. Decentralized protocols can issue tokens to pay for compute, diluting existing holders but avoiding the depreciation hit that Tencent faces. However, token-based funding introduces volatility and governance complexity. During my 2022 bear market research, I spent 200 hours modeling recursive SNARKs for a ZK-rollup visualization tool. I learned that capital efficiency in crypto is not about dollars spent, but about aligning incentives. Tencent’s capex is a top-down decision driven by a central strategy. In decentralized AI, capex is distributed across thousands of node operators, each making individual decisions based on token rewards. This is the difference between a planned economy and a market economy.
Take the example of Bittensor. Its subnet architecture allows anyone to offer compute or training data, earning TAO tokens. The total value locked in Bittensor’s subnets is around $500 million—a tiny fraction of Tencent’s HKD 215.7 billion. But the capital efficiency is different. Bittensor’s network produces intelligence without a centralized depreciation schedule. Participants bear their own hardware costs, and the network only pays for useful work. This is analogous to how Curve Finance’s stableswap invariant minimized impermanent loss compared to traditional market makers. The same principle applies: decentralized systems can be more capital-efficient because they distribute risk. However, they also suffer from coordination failures. Tencent can decide to build a single massive data center in Guizhou tomorrow. Bittensor requires hundreds of miners to agree on a subnet upgrade, which takes weeks of governance. The bear market didn’t change this fundamental trade-off; it just made it more visible.
Now, consider the specific data from the CITIC report. Core profit growth of 19% shows that Tencent’s core businesses (gaming, ads, fintech) are healthy enough to subsidize AI. The report notes that “profitability release of core businesses empowered by AI supports the investment in AI.” In other words, AI is already generating returns through better ad targeting and game content creation. This is a virtuous cycle. In decentralized AI, we haven’t seen a similar loop yet. Most AI crypto projects are still in the infrastructure phase—building compute markets, data registries, or inference protocols. The killer app that generates recurring revenue on-chain hasn’t emerged. I experienced this firsthand with TruthLayer. In 2025, my team built a decentralized registry for AI-generated media, integrating watermarking with IPFS. We got 500 beta testers in a month, but zero revenue. Users loved the narrative of “human oversight,” but they weren’t willing to pay for it. The decentralized AI ecosystem is still searching for its “advertising” moment—a product that generates such clear value that it funds the entire stack.
Contrarian: The Blind Spot of Centralized AI Capex The conventional wisdom is that Tencent’s massive capex will crush decentralized competitors. But that’s a surface-level reading. The bear market didn’t teach us that big money wins; it taught us that resilience comes from adaptability. Tencent’s depreciation costs are a hidden vulnerability. If the AI boom slows or if a new model architecture (like a more efficient transformer variant) reduces compute requirements, Tencent’s data centers become stranded assets. In crypto, node operators can simply turn off their GPUs and sell them. There’s no corporate balance sheet to manage. This flexibility is a form of antifragility. Moreover, Tencent’s AI strategy is fundamentally centralized. It controls the models, the data, and the distribution. This creates regulatory risks—governments could demand censorship of certain AI outputs, or impose data localization rules. Decentralized AI, by design, is harder to censor. A model running on Bittensor or on a smart contract that verifies inference with zero-knowledge proofs can’t be shut down by a single authority. This is a values-driven advantage that no amount of capex can replicate.
Another blind spot: Tencent’s capex is based on the assumption that AI will remain a winner-take-all market. But the history of the internet suggests otherwise. The early web saw massive centralized investments in AOL and CompuServe, yet the open protocols (HTTP, SMTP, TCP/IP) ultimately won. The same could happen with AI. The open-source model movement (Llama, Mistral, etc.) is already challenging proprietary models. In crypto, we have the potential to build open protocols for AI training and inference that are owned by the community. The capital expenditure would then be distributed across thousands of participants, each contributing small amounts. This is more resilient. During the 2022 crash, I saw how centralized exchanges collapsed while decentralized exchanges like Uniswap kept functioning. The same principle applies to AI infrastructure. Centralized compute pools are honeypots for hackers and regulators. Decentralized compute markets are harder to attack.
Furthermore, the CITIC report highlights that Tencent’s core net profit growth will slow to 2-3% in 2026-2027 due to depreciation. Meanwhile, the company is cutting stock buybacks and dividends to fund capex. This financial pressure is real. In decentralized networks, there is no such pressure to return profits to shareholders. The protocol can reinvest all value into growth or distribution. This is why I believe that the real AI winner will be a hybrid model—a decentralized coordination layer that taps into centralized compute resources when needed, but ensures sovereignty through cryptographic proofs. Projects like Ritual or Gensyn are exploring this path. They don’t compete with Tencent on scale; they compete on trust and composability.
Takeaway: The Future of Intelligent Infrastructure The bear market didn’t kill the dream of decentralized AI; it gave us the clarity to see where the real value lies. Tencent’s HKD 215.7 billion capex is a vote of confidence in the importance of AI, but it’s also a warning about centralization. As a blockchain builder, I see an opportunity. We don’t need to match Tencent dollar-for-dollar. We need to build protocols that are more capital-efficient, more resilient, and more aligned with human values. The next phase of AI won’t be about who has the most GPUs; it will be about who has the most trustworthy network. The 2026 capex figures from CITIC are a benchmark, not a barrier. They tell us that the compute demand is real. Now it’s up to the decentralized community to channel that demand into permissionless infrastructure. About me: I’m Chris Thompson, a 29-year-old protocol PM in Nairobi who started coding in 2017 to understand the DAO hack. I’ve seen cycles come and go. The bear market didn’t break my spirit; it focused my mission. And I believe that the marriage of AI and blockchain is the most important frontier of the next decade. The numbers from Tencent just prove that the prize is worth fighting for. The question is: will we build the open protocols before the moats become too deep?