Dolphin Research

Zhipu 1H26 First Take: against a ~$70bn market cap, 1H revenue was $950mn, which is hardly exciting. GPM for the open platform and API remains low at 25% despite the sharp revenue scale-up, still below 30%.

What stood out vs. expectations was elsewhere. The surprises were:

a) Revenue mix shifted decisively, with the open platform and API now 86% of revenue. This largely removes concerns over Zhipu's much-disliked bespoke model project business. The market expected this mix to rise, just not by this much.

b) Spend was highly disciplined, with sales and admin down. R&D came in at just RMB 2.1bn, well below market expectations.

That said, the lengthy filing pales in comparison to a short headline: Aug ARR ran at $1.6bn with GLM 5.3 and 5.3 Flash, vs $1.0bn in Jul. A $1.6bn ARR can paper over any blemishes in the report. So Zhippu can keep dancing!$Z.AI(02513.HK)

Horizon Robotics 1H26 First Take. The company had pre-announced a profit warning, and 1H revenue and blended GPM landed toward the high end of guidance; however, guidance had already been cut vs. prior Street consensus, and the upside was driven by licensing and services.The core product solutions biz (chips + HSD algorithms) materially missed. Both revenue growth and margins slowed sharply; breakdown below:

1) Product solutions revenue slowed notably. The 'volume + price up' thesis remains unproven.

Revenue growth decelerated sharply: Q2 product solutions revenue was just RMB 930 mn, with growth slowing from 85% in 2H25 to only 15% in 1H26. This is far below the prior 2026 full-year guidance for the segment of +132%.

Shipments growth eased: 1H26 chip shipments were only 2.218 mn units, up just 12% YoY and below the full-year guide (Horizon has cut its FY shipments guide from 5.42 mn sets to 5.00 mn, implying ~25% YoY). The weakness reflects soft domestic PV retail demand in 1H26 (-20.2% YoY) and sales at key customer BYD Co. missing expectations, while BYD’s in-house high-compute 'Xuanji A3' chip (3-chip cluster >2,100 TOPS) stoked concerns about share loss in intelligent-driving chips.ASP gains stalled: 1H26 ASP was RMB 417, up only 2.4% YoY, indicating the higher-priced J6P has yet to scale. Shipments remain concentrated in the Mono series (J2, J3, J6B) and other mid/low-tier chips. The market’s thesis of domestic substitution led by the high-compute J6P driving both volume and ASP higher has not materialized in 1H26.

2) Product solutions GPM fell YoY, pressured by three factors.

Q2 product solutions GPM was just 36.2%, down 600 bps YoY and well below the 46.7% expected. Key drags included:

a) Product mix skewed to mid/low end. Shipments remained dominated by the lower-priced Mono series, while the high-ASP, high-margin J6P contributed minimally.

b) Memory price upcycle eroded hardware profits. DRAM prices continued to rise, and higher memory costs directly compressed product margins.

c) Higher mix of low-margin system-level deliveries. To accelerate roll-out of high-end HSD (SuperDrive) solutions, Horizon provided system-level deliveries including domain controllers to select clients. The low-margin domain-controller hardware diluted overall solution margins.$HORIZONROBOT-W(09660.HK)

Birentech 1H26 quick take: Post-market on Aug 28, Birentech released its first results since listing. As the company had already issued an interim profit alert on Aug 17 with ranges for revenue and loss, the better-than-expected narrowing was largely priced in.However, the next-day and subsequent tape suggested investors were not buying it.

In our initiation, we laid out the stock’s beta logic and H2 catalysts. This note tracks what in the print merits attention.

1. Revenue

Despite revenue surging +2,000% YoY to RMB 1.236bn, QoQ growth was only +26.6%, indicating no sustained step-up. The core revenue driver remains BR166X, which we flagged as a dated product in our coverage.Market focus is still squarely on BR20X.

2. Gross margin

GPM expanded 1,100bps YoY, yet it is clearly lower vs. peers Moore Threads/MetaX with near-60% GPM. This is largely due to the current mix being BR166X-heavy.Mass production of BR20X should improve margin structure.

3. Leading indicators

Downstream CSPs such as Tencent and Alibaba have materially ramped CAPEX, but this cycle has yet to show up in Birentech’s leading indicators. Inventories rose 28% vs. FY25, broadly in line with revenue growth, while contract liabilities shrank slightly.This suggests downstream orders have not flowed in meaningfully to Birentech, as BR166X lags peers by a generation and has not broken through with customers.

BR20X has taped out in small volumes, and prepayments rose +230% vs. end-FY25, likely reflecting upstream capacity reservations. We expect a meaningful ramp for BR20X in 2027.

4. Valuation

The company’s current market cap is RMB 87bn. We see domestic compute as a clear investment theme, with BR20X commercialization and customer breakthroughs remaining the key H2 watchpoints.That said, the fundamentals in this print do not point to a strong alpha name.

We think the tape-out validation cadence is not fast, and BR20X, initially slated for H2 commercialization, may be delayed. In the supply chain, leading foundry SMIC’s advanced-node capacity is primarily allocated to Huawei and Cambricon, and Hua Hong’s capacity ramp is another key risk.

In Dolphin Research’s TAM sizing, the AI xPU market reaches $80–90bn by 2030. Under a conservative case of 1.5% share (our base-case market split: Huawei ~70%, Cambricon ~10%, and if the remaining eight players split evenly, Birentech at ~2.5%), that implies revenue of $1.2–1.35bn.Assuming ~15x PS in 2030 and a 40% margin (≈38x PE), discounted back to 2026 at 13%, the fair value is $11–12bn, or ~HK$80bn.We see better downside protection at that level. $BIREN TECH(06082.HK)

Meituan Q2 2026 First Take: At first glance, results looked strong. Revenue came in slightly above expectations, with a bigger surprise on profits vs. Bloomberg consensus.Both core local commerce and group OP beat by over RMB 2bn. Initial take: profit recovery is tracking better than expected.

On the one hand, Dolphin Research believes Meituan is firmly on a recovery path. With the food-delivery price war fading, both growth and profitability are improving.The local commerce segment also swung to profit this quarter, and not by a small margin.

However, we would caution that the headline profit lacks quality. The quarter booked over RMB 3bn in other gains from shutting down Select and other investments.Excluding this, group OP was still hovering around breakeven.

Also, while most metrics beat Bloomberg consensus, they were merely in line with top-tier houses such as Goldman Sachs. In short, the recovery is real, but not as stunning as it first appears.$MEITUAN(03690.HK)

XD Inc. 1H26 quick take: results came in below Street expectations, as the biz hit a trough in the game release cycle and stepped up TapTap AI spending. With revenue growth slowing, core margins also slipped slightly.

H2 should see a natural recovery on an easier base and continued international rollout of 'Ragnarok Beginner's Server' in Japan/Korea. A return to strong growth still hinges on new titles; with three projects in the pipeline, the next product cycle is likely to start next year.

1) Game growth slowed; next cycle starts next year

Game revenue reached RMB 2.3bn in 1H26, up 12% YoY, but growth slowed QoQ off last year's high base. Contributors were the international performance of 'XD Town' and resilient 'Torchlight' as the floor. New titles included 'Resonance of the Star Mark' launched last Dec and premium title 'Dave the Diver'.

On the positive side, player counts and ARPPU were stable to slightly higher, indicating limited core user churn. Deferred revenue was RMB 290mn at end-1H; implied billings grew 16%, outpacing recognized revenue.

'Ragnarok Beginner's Server' may continue to expand internationally in H2 as the base eases. The key watch is 'Ragnarok RO 守爱 2', with performance after next year's open beta.

2) Ad revenue declined: lack of new hits, deliberate traffic allocation, and a high base

TapTap ad revenue was under RMB 1bn in 1H26. It fell 1% YoY.

On one hand, last year's base was elevated by several casual-game hits that spent heavily on UA, and XD had no blockbuster domestic launches in 1H to pull TapTap traffic, weighing on overall exposure. This also dampened platform-wide traffic.

On the other hand, XD rolled out TapTap Maker, a conversational AI game-creation agent. Over 6,000 games have been created via Maker and published on TapTap, receiving some internal traffic tilt, which affected TapTap's monetization.

3) Profit under dual pressure, down YoY

As net profit includes non-recurring items unrelated to core ops, we focus on core operating profit (GP minus SG&A/R&D). This better reflects underlying performance.

With top-line growth slowing, AI-related R&D in TapTap Maker and 'Cindy' weighed on profit, alongside incremental amortization from acquiring the 'Torchlight' IP. Core operating profit fell 6.2% YoY in 1H26; margin decreased by 2pct YoY and 5pct QoQ.

4) Most negatives are priced in

The current market cap is ~HKD 21bn (RMB ~18bn). On our FY26 core operating profit estimate of RMB ~1.8bn (below pre-report Street consensus of ~RMB 2.0bn) and an effective tax rate of 15%, P/E is under 12x, implying limited TapTap premium vs. game peers. Further downside looks limited; trading opportunities likely emerge in Q4 or near year-end as the new cycle approaches, supporting sentiment. $XD INC(02400.HK)

Marvell First Take: Results broadly in line with market expectations, with growth on the top line driven mainly by data center interconnect products.

Starting this fiscal year, the company consolidated segment disclosures from five to two: Data Center, and Comm & Other.

With a full-year outlook provided, next-quarter guidance matters less.

More importantly, management gave a two-year revenue outlook.

Dolphin Research believes Marvell’s post-market drop is mainly due to two factors.

The stock had traded at a premium to NVDA and peers, reflecting expectations for share gains in AI chips.

Bilibili 2Q26 First Take: Q2 results were mediocre, broadly in line with expectations. Revenue was again carried by ads, with growth ticking up QoQ on easier comps. On the bottom line, a slight beat came mainly from tighter control of selling expenses.

This year’s key issue is a thin games slate. The high base from 'San Mou' drove a sharp YoY decline in gaming revenue. While the in-house title 'Baijiangpai' launched in Q2, its scale is too small; to move the needle, they need at least a mid-sized breakout.

In H2, the revenue comp base falls quickly, which should naturally ease the trend. Based on the current pipeline, two in-house titles 'Shine, Lumi!' and 'Three Kingdoms: The Way' (many expect both could be bn-level) are slated for Q4 launch. The new-release cycle could revive gaming growth.$Bilibili(BILI.US) $BILIBILI-W(09626.HK)

MXBC 1H26 quick take: results missed clearly, with revenue almost flat at +2.3%, far below JP Morgan’s +11% preview in Jul. Profit fell 14.7% YoY, marking the first half-year profit decline since listing. Selling expenses rose 23% YoY and admin expenses 39.4% YoY; with revenue up only 2.3%, the combined expense ratio expanded by 230bps.

In detail. Key points below.

1) The biggest issue is that revenue has effectively stalled. 1H26 revenue was RMB 15.22 bn, up just 2.3% YoY, and after delivery subsidies faded, growth decelerated from around 30% to near zero within half a year.

2) Store openings beat, but same-store performance deteriorated sharply. Franchise stores net-added 4,166 in the period (5,455 openings, 1,289 closures), taking the global network to about 64,000 stores by period-end, up 20.7% YoY, while the closure rate improved from 2.55% to 2.16%.

In other words, the feared slowdown in new openings did not materialize; the real deterioration is at the unit level. Note that mix effects from lower per-store scale at Xingyunka and Fresh Beer Fulujia contribute to this, but the drop remains sizable beyond mix.

3) Flat revenue, but much higher opex. Selling and distribution expenses were RMB 1.12 bn (+23% YoY), lifting the ratio from 6.1% to 7.4%, and admin expenses were RMB 610 mn (+39.4% YoY), with the ratio up from 2.9% to 4.0%.

4) On shareholder returns, the BOD proposed a special dividend of RMB 2.65 per share around Nov. 6, totaling about RMB 1.01 bn, implying a ~43% payout of 1H net profit but only ~1.25% one-off yield vs. the current market cap, offering limited support.

Dolphin Research previously preferred entry below 15x; the multiple now looks close, but the profit base has been cut from the prior ~RMB 6.7 bn estimate to ~RMB 5.2 bn, meaning both price and earnings are falling, and the multiple has not really de-rated.