---
title: "AI Cloud + Kunlun Chips: Time to Re-rate BIDU?"
type: "Topics"
locale: "en"
url: "https://longbridge.com/en/dolphin/post/43787192.md"
description: "After so many years and so many calls, talk of a re-rating for $Baidu(BIDU.US) convinces few. Relying on the legacy internet biz or on autonomous driving won't deliver a re-rating.Re-rating $BIDU-W(09888.HK) means breaking the 'old' businesses, treating legacy ads as a cash-flow unit with zero valuation. Under that premise, reassess, in the AI era, the value of BIDU as an AI infrastructure stock. The 'AI infra' version of BIDU, in Dolphin Research's view..."
datetime: "2026-09-07T12:19:35.000Z"
locales:
  - [en](https://longbridge.com/en/dolphin/post/43787192.md)
  - [zh-CN](https://longbridge.com/zh-CN/dolphin/post/43787192.md)
  - [zh-HK](https://longbridge.com/zh-HK/dolphin/post/43787192.md)
author: "[Dolphin Research](https://longbridge.com/en/dolphin.md)"
generator: "portal-rs"
---

# AI Cloud + Kunlun Chips: Time to Re-rate BIDU?

After so many years and so many calls, talk of a $Baidu(BIDU.US) re-rating barely moves the needle anymore. Relying on legacy internet or Robotaxi will not deliver that re-rating.

Reframing $BIDU-W(09888.HK) means breaking up the 'old' biz — treating the legacy ads unit as a cash-flow engine with zero equity value. With that baseline, reassess Baidu as an AI infrastructure play in the AI era.

In our view, 'AI infra' at Baidu rests on two core assets. These are the key to any re-rating narrative.

**1) AI cloud services:** Bare Metal as a Service, and MaaS focused on API distribution.

**2) Kunlun Chip:** ASIC-based compute focused on selling chips and rack-level solutions as a chip design service.

We pulled Baidu out for a fresh look to weigh these two assets and check for real value. We go through them one by one. 

**I. AI cloud: start with the 3-year payback model used by hyperscalers**

In Q2, both US CSPs like Amazon and China CSPs like Alibaba and Baidu gave a quantified 3-year payback timeline for AI investments. The goal is to ease investor concerns about elevated near-term Capex.

There is an implicit assumption that unit token GPM stays relatively stable. Software-side iteration (FlashAttention, speculative decoding, MoE, prefix caching, etc.) has boosted effective chip utilization, unlocking more compute throughput.

As a result, while per-token pricing is falling fast, per-token cost is also dropping. For older GPUs in particular (implying limited single-card price hikes), unit token cost may fall faster than price, which can stabilize or even lift token-level margins in the near term. 

**The question is sustainability, which remains uncertain.** Hardware utilization has a ceiling and requires increasingly complex software optimization to approach it. Open-source models and price competition are compressing token revenue faster.

As the chart shows with H100 cloud rentals, 3-year rental prices are down 78% while single-card throughput is up 8x, driving a 97% drop in per-token cost. Token GPM rose to ~75% in 2024, then stopped improving. In H1 this year, with further end-price cuts for LLMs and utilization ceilings on older cards, token GPM even showed signs of softening.

**Compute cost is still the main hurdle to wider AI adoption.** Extending useful life is now common, and Jensen Huang personally endorsed A100 use through 2029, implying a 10-year life.

But the chart also shows life extension on old cards is a stopgap. H100 can no longer defend unit economics in a price war, and whether A100 truly fits LLMs for 10 years is hard to promise. **Industry resources should still tilt toward easing hardware supply bottlenecks.**

**1\. US vs. China CSP payback gaps create room for domestic chips**

This AI cycle has five main monetization paths for CSPs, with current growth and positive ROIC concentrated in the first two (IaaS and MaaS). The mix matters for returns.

Recent sell-side models suggest current compute pricing and costs can support a 3-year payback. North America runs higher margins than China, and MaaS runs higher margins than other lines. A key reason: 1P model R&D is not fully burdened to MaaS, while 3P model distribution is channel monetization recognized on a net commission basis, which lifts margins.

A rough valuation take:

(1) For US IaaS with self-built DCs, ROIC can reach ~30% (GB300 as an example, rental revenue $23.0bn/yr vs. upfront Capex $39.0bn/yr). On OCF, the cash payback is ~2.2 years. 

(2) For US MaaS offering multiple model APIs (1P and 3P), compute can be self-built or rented from third parties. Excluding 1P training costs, marginal OP margin can be ~75% (self-built compute) and ~30% (rented compute). **ROIC is ~46% for self-built, with payback under 2 years; rented compute has no upfront Capex.**

That said, the split with 3P model providers matters, and US CSPs lack pricing power vs. Anthropic and OpenAI. For example, Amazon Bedrock is a distribution platform, and with weaker in-house Nova, it primarily sells models from ~20 AI labs globally, including China.

Q2 AWS growth re-accelerated on Anthropic’s strength (channel checks suggest Bedrock is Anthropic’s largest CSP channel, contributing about half of its revenue). Financially, Bedrock books net revenue, while **Anthropic books gross revenue.** Anthropic deducts inference costs from gross, then splits the remainder 50/50 with Bedrock as channel fees, which is largely profit accretive to Bedrock and structurally higher-margin than IaaS.

(3) For China CSPs, unit compute cost advantages exist in mid/low-end accelerator cards, O&M, and bundled software. For H200-class and above, export controls drove extreme price volatility. Chinese CSPs often pay multi-fold premiums to procure GPUs, but cannot pass that through on rental pricing and must consider client affordability, sometimes even discounting vs. US peers, compressing GPM.

Despite higher US O&M (power, rack space, network, labor, etc.), lower operating margins and higher GPU procurement costs in China drive roughly half the ROIC vs. US IaaS peers. 

On Alibaba’s call, management argued Morgan Stanley’s ROIC was understated, implying ROIC north of 20%. The main discrepancy likely comes from compute chip cost treatment.

Morgan Stanley simplified by assuming all compute is GB300, but US and China CSPs both run hybrid stacks. Inference does not always need the highest-end cards, often mixing A100/H100-class or domestic chips, which do not carry 3–4x premiums in China, lifting realized GPM after depreciation. 

As domestic chip performance improves (higher single-card FLOPS, large-scale networking to offset per-card gaps) and total cost advantages emerge at system level (more cards, higher power draw, rack space, plus servers, routers, switches), and with overseas procurement compliance risk, China CSPs began partial switching this year. In short, **at least on inference, domestic chips now hold a relative cost edge, creating real shipment opportunities.**

**II. Compute windfall: Baidu’s late-cycle ‘second spring’**

Baidu now has Kunlun plus external chip sourcing, and to a degree is enjoying a full-stack AI infra tailwind in AI cloud. That positioning matters for the next leg.

Baidu highlighted its ‘AI biz architecture’ from Q3 last year, but the compute windfall started in late 2023, with GenAI cloud revenue truly scaling in H2 2024 (10%+ of cloud revenue). The AI stack splits into AI cloud infra, AI apps, and AI-native marketing.

**The windfall sits mainly in AI cloud infra** — comprised of IaaS, MaaS, and Kunlun (external sales). IaaS is the primary revenue driver, while the higher-margin MaaS channel model is smaller at Baidu vs. Alibaba and ByteDance due to weaker distribution.

Baidu AI cloud’s edge: First, on supply, it has available GPU capacity and mature public-cloud services, as capacity prepared for Ernie can be leased out, and Kunlun procurement eases building AI cloud. Second, on demand, Baidu’s neutrality vs. Alibaba/ByteDance helps attract PDD in e-comm, Kuaishou in short video, and major game studios like HoYoverse, plus gov./SOE clients in power/energy.

Hence the company’s repeated disclosure of rising GPU cloud subscription revenue, which has accelerated QoQ over the past year, signaling robust demand. Momentum has been building steadily.

Baidu currently has 1GW of compute and expects to double in 1–2 years. It will likely use flexible financing such as leasing rather than pure chip Capex to lower near-term cash needs and support ROIC. 

In the near to mid term, modest capacity expansion during the compute windfall should support cloud growth. Scale with discipline is the focus. 

Q2 Capex spiked, raising concerns about aggressive investment. On earnings calls, both Baidu and Alibaba discussed an AI ROIC model that implies a reference point — **'30%+ revenue CAGR + 30% GPM'** can deliver payback within 3 years.

**(1) Steady-state margin view:** Signed orders need at least 30% GPM. Deduct 15–20% for steady Opex, and OPM can reach ~15%, a sustainable level for operations. 

**(2) Cash payback view:** Each unit of self-built compute uses a 5-year depreciation life. With 30–40% GPM on RMB 10bn revenue, depreciation would be RMB 6.5bn, implying upfront Capex of 6.5×5 = RMB 32.5bn. At zero growth, Opex cash outflow of RMB 2bn would leave OCF of RMB 8bn, implying >4 years to recover RMB 32.5bn.

With revenue growth, payback accelerates. Sensitivity suggests that with 30% GPM, only 30%+ revenue CAGR gets payback within 3 years, and lower GPM requires even higher growth to match that timeline. 

**The sensitivity shows that stabilizing GPM eases payback pressure versus relying on growth.** GPM tracks client pricing and near-term Capex cadence; pricing is mostly market/contract-driven, but Capex timing is controllable.

Management noted in a callback that Q2 Capex was a procurement peak and will not keep making new highs. **Baidu aims for quality growth in cloud, not scale at any cost, to avoid impulsive spending.** After 1–2 years of concentrated investment, payback and cash generation can support the group’s earnings more meaningfully.

On a cautious note, this also suggests management is not fully confident in the durability of today’s elevated compute spreads. **They prefer locking 3–5Y LT contracts to secure pricing certainty during the windfall.** As supply loosens and competition intensifies, incremental demand will push for lower pricing, and compute spreads will fade at the margin.

**III. ASIC value: order boom vs. capacity choke — can Kunlun still fetch $50bn?**

Baidu is arguably the first in China to form a compact AI full-stack, but versus Google’s 2024 stack it is a lite version. The gap is not only in models but also in the chip race within its peer set, weakening the flywheel.

Based on performance design and actual mass production (shipments), domestic chips fall into three tiers. The pecking order today is as follows.

Tier 1 is **Huawei Ascend**, which leads on single-card performance and networking, has adapted to top LLMs, and most importantly enjoys priority allocation at advanced nodes. That capacity edge is decisive.

Tier 2 includes **Cambricon, Hygon, Pingtouge**, and **Kunlun**, each with strengths. Paper specs no longer create absolute moats across designs; the key gap is realizing volume capacity.

Tier 3 includes **Moore Threads, Biren, Denglin, Iluvatar**, and **Metax** among others, with Metax relatively stronger on performance and shipments. Aside from Iluvatar, which is near IPO, most are already listed to fund R&D.

Kunlun is not capital constrained given BIDU’s ample cash. However, within Tier 2, the first three have migrated from overseas to domestic foundries, while **Kunlun’s current mainline P800 still relies on Samsung capacity.**

The new M100 has only taped out for client tests and is not yet in mass production, while M300 is in design targeting shipments in H2 next year. With the Samsung framework expiring next year and sanctions risk, **a switch to domestic foundries is inevitable, and ramp timing and scale are the biggest uncertainties for the M series.**

Domestic process migration requires architectural tweaks, which take time, so the schedule could slip as it did earlier due to capacity. **Even with a strong order book (we estimate near RMB 10bn),** capacity constraints will affect revenue realization cadence and weigh on valuation under a PS-plus-peer-comps framework.

Market chatter in late May suggested an IPO range of $40–50bn for Kunlun. **A back-of-the-envelope math likely assumes 2026 revenue of RMB 8bn (roughly 150k P800 units) and a 40x P/S.**

Given foundry migration risks, we haircut realization: **use 2024E revenue of RMB 8bn (vs. RMB 3.5bn last year, +120% YoY) and a 30x P/S, implying ~$35bn.** BIDU owns just under 60% today; after IPO dilution we assume a conservative 55% control.

Adjusting for related-party sales — BIDU contribution fell from ~60% last year to below 50% this year and may stabilize near 30% — we estimate the external portion and apply the 55% stake. **That yields roughly $13.6bn attributable to BIDU.**

BIDU’s current market cap is ~$33.0bn. Netting out Kunlun equity value, the parent is below $20bn, almost equal to net cash (cash + ST investments − ST borrowings), similar to Kuaishou vs. its Keling stake plus net cash.

**While legacy search has little upside, we see valuation as depressed (vs. our neutral fair value of ~$39bn, with ~20% upside to repair), and positioning remains cautious.** Notably, southbound inclusion today coincided with a 5% drop, a classic sell-the-news.

With Kunlun’s IPO ahead, BIDU will likely prioritize credit over equity financing. Coupled with the $5bn 3-year buyback, downside looks limited, but timing depends on Kunlun’s listing progress and capacity milestones. Given carry costs, we prefer Q4 into year-end — near Kunlun’s listing window and potential M100 mass deliveries — to play a possible **rebound**.

\<End here>

**Dolphin Research on 'Baidu' — archive:**

**Latest earnings season**

May 18, 2026 Transcript: [Baidu (Trans): Full-stack AI advantages lie in cost efficiency and deployment stability](https://longbridge.cn/topics/40885468?channel=SH000001&invite-code=355628&app_id=longbridge&utm_source=longbridge_app_share&locale=zh-CN&share_track_id=f3014d74-3683-4b95-a7bb-6f441db9885a)

May 18, 2026 Earnings take: [Baidu: Little else to lean on — all about Kunlun](https://longbridge.cn/topics/40881952?channel=SH000001&invite-code=355628&app_id=longbridge&utm_source=longbridge_app_share&locale=zh-CN&share_track_id=86283cb1-5edd-4107-bdd1-3ca02b7705ff)

Feb 26, 2026 Transcript: [Baidu (Trans): Sharper focus on shareholder returns and execution](https://longbridge.cn/topics/38917013?channel=SH000001&invite-code=355628&app_id=longbridge&utm_source=longbridge_app_share&locale=zh-CN&share_track_id=72637625-3b78-4806-94ff-992efd680880)

Feb 26, 2026 Earnings take: [Baidu: Ads base eroded — can Kunlun be the lifeline?](https://longbridge.cn/topics/38912994?channel=SH000001&invite-code=355628&app_id=longbridge&utm_source=longbridge_app_share&locale=zh-CN&share_track_id=fa24cd37-32b5-4dba-9317-e6a3cdf399ba)

Nov 18, 2025 Transcript: [Baidu (Trans): AI-native marketing as the second growth curve](https://longbridge.cn/topics/36483983?channel=SH000001&invite-code=355628&app_id=longbridge&utm_source=longbridge_app_share&locale=zh-CN&share_track_id=d9f048be-6399-45eb-9317-45a2f8c61d52)

Nov 18, 2025 Earnings take: [Heavy AI storytelling — is Baidu really fixable?](https://longbridge.cn/topics/36480554?channel=SH000001&invite-code=355628&app_id=longbridge&utm_source=longbridge_app_share&locale=zh-CN&share_track_id=03597d68-a7bf-460a-8e85-59a2b10f94f5)

**Risk disclosure and statement:** [**Dolphin Research disclaimer and general disclosures**](https://support.longbridge.global/topics/misc/dolphin-disclaimer)

### Related Stocks

- [K3SD.SG](https://longbridge.com/en/quote/K3SD.SG.md)
- [09888.HK](https://longbridge.com/en/quote/09888.HK.md)
- [89888.HK](https://longbridge.com/en/quote/89888.HK.md)
- [BIDU.US](https://longbridge.com/en/quote/BIDU.US.md)
- [09988.HK](https://longbridge.com/en/quote/09988.HK.md)
- [BABA.US](https://longbridge.com/en/quote/BABA.US.md)
- [89988.HK](https://longbridge.com/en/quote/89988.HK.md)
- [HBUD.SG](https://longbridge.com/en/quote/HBUD.SG.md)
- [HBBD.SG](https://longbridge.com/en/quote/HBBD.SG.md)

---
> **Disclaimer: This article is for reference only and does not constitute any investment advice.**