AI Weekly Malaysia

Summaries

Short AI and tech summaries with source links, signal scores, and why each update matters for builders, founders, and Malaysian tech workers.

Reset

Showing 1-23 of 23 results

DateProviderScoreSummary
18 Nov 2026, 7:37 AMTechCrunch8.5 Nvidia is abandoning China after Jensen Huang’s ouster

Nvidia acting CEO Colette Kress said the company will stop selling H200 chips in China, ending a Huang-brokered arrangement that routed shipments through the U.S. with a 25% tariff. The disclosure came on Nvidia’s first earnings call since Jensen Huang stepped down, alongside new U.S. rules barring advanced chips in foreign data centers from serving Chinese customers remotely. Jefferies estimates those rules could cut $44B–$62B from Nvidia’s data center revenue next year; Kress said Vera Rubin backlog means no top-line impact, while the company flagged possible long-term customer losses in Singapore, Malaysia, Thailand, and Indonesia.

Why: If you rent GPU capacity in Southeast Asia or plan AI infrastructure, ask your provider whether it can serve Chinese-linked workloads remotely and what U.S. export-license constraints apply. The article specifically names Malaysia among markets where Nvidia may lose potential customers, and says Nvidia’s Singapore subsidiary hit $104B in revenue in 2025—so local data-center and GPU-cloud plans should model tighter compliance and possible capacity shifts, not assume unchanged H200-class supply.

19 Nov 2026, 2:53 AMHugging Face Blog7.5 Why your fine-tuning job takes longer than the FLOPs say

A Dharma-AI team analysed 272 real supervised fine-tuning jobs from their production pool of NVIDIA H200 GPUs (~7,382 GPU-hours) to test how well the standard FLOP-based time estimate holds up. The closed-form 6·P·D / (N_GPU · FLOPS_peak · MFU) calculation explained little of the variance (R² = 0.31, MAE 12.41 h), while a CatBoost regressor using only pre-dispatch information reached R² = 0.98 with MAE 1.71 h and median error 0.51 h in 5-fold cross-validation. The gain comes from signals a FLOP formula cannot see: evaluation cadence, checkpoint policy, per-GPU load, and fine-tuning regime, plus the nonlinear interactions between them. This is a preview of their forthcoming paper, 'Predicting End-to-End Fine-Tuning Time for Generative Language Models'.

Why: If you plan GPU spend or queue order with a FLOPs/MFU back-of-envelope, this dataset says your estimate is off by roughly 12 hours on average — enough to leave H200s idle or push every queued experiment back. The concrete fix is to start logging four things at job submission: evaluation cadence, checkpoint policy, per-GPU load at dispatch, and fine-tuning regime — those are the inputs that got the regressor to ~1.7 h MAE, and you cannot train that model later without them.

16 Nov 2026, 11:02 PMTechCrunch7.5 US tightens AI chip export controls to deter Chinese companies from accessing compute remotely

The US Commerce Department is rolling out an interim rule requiring licences to export, reexport or transfer AI chips and servers to data centres anywhere in the world if those facilities may supply compute to companies in the D:5 embargoed group — China, Russia, Iran, North Korea and 18 others. The rule kicks in at data centres or networks capable of more than 12 million total processing performance (TPP), which the article says is roughly the max compute of 750 Nvidia H100 or H200 chips. It closes the remote-access loophole rather than the physical shipment one, and follows Reuters reporting that Chinese AI firms are rushing to lock in overseas compute capacity.

Why: The licence trigger is customer reach, not geography — a data centre in Malaysia, Singapore or anywhere else needs a US licence if it may serve a D:5-based customer at 12M+ TPP scale. If you rent GPU capacity from a regional provider, or you resell/host GPU boxes for overseas clients, ask your provider now whether their ownership chain and customer list could put them inside this rule; a provider that gets caught can have your capacity cut off with little notice. Founders building on overseas GPU clouds should also check contract terms for compute-supply disruption, since the article notes Chinese buyers are already absorbing capacity.

18 Nov 2026, 3:16 PMCNBC Technology6.5 Nvidia warns chip buyers of legal risks tied to China AI access, sources tell CNBC

Nvidia warned customers that the proposed Remote Access Security Act could expose them and Nvidia to U.S. legal action over Chinese remote access to AI chip capacity, according to CNBC sources. Nvidia also said it would not sell chips to customers that rent more than 50% of their GPUs to Chinese companies. Current U.S. controls already restrict advanced chip shipments to China, while Congress is considering broader authority over chips hosted in overseas data centers.

Why: If you buy or resell cloud GPU capacity, the 50% Chinese-tenant cutoff and RASA's possible extension to remote cloud access are concrete procurement and compliance risks. Ask providers how they count tenants and what happens to your contract if the law passes, especially if you rely on overseas GPU clouds.

18 Nov 2026, 2:21 PMCNBC Technology6.5 'Who and what comes next': Analysts expect Nvidia investor uncertainty despite earnings beat

Nvidia reported third-quarter revenue of $115.84 billion versus $109.10 billion expected and guided fourth-quarter revenue to $130 billion versus $125.1 billion consensus, with Melius Research estimating an extra fiscal week accounts for $9 billion to $10 billion of the sequential jump. Nvidia also raised its fiscal 2028 growth forecast to 75% from 70% while excluding China from its data-center compute revenue outlook. Analysts still expect investor volatility and uncertainty over 'who and what comes next' after Jensen Huang's resignation as CEO, even as shares rose 5% after hours.

Why: Builders planning GPU-dependent AI workloads or SaaS products should treat the $130 billion Q4 guide and 75% fiscal 2028 growth target as a signal that Nvidia expects AI compute demand to remain strong, so assuming GPU prices will fall soon may be risky. The explicit exclusion of China from the data-center compute outlook also means export and regional supply constraints should stay in capacity planning.

18 Nov 2026, 5:48 AMTechCrunch6.0 SEC reportedly investigating Nvidia amid Huang exit

The Wall Street Journal reports the SEC has opened a civil probe into whether Nvidia's disclosures misled investors, arriving shortly after CEO Jensen Huang resigned following disclosures of extramarital affairs, including one with Alice Huang, an executive at Singapore-based data company Megaspeed. The Commerce Department has been investigating Megaspeed since late 2024 over suspected diversion of Nvidia chips to China, and Treasury Secretary Scott Bessent publicly accused Nvidia of knowingly selling powerful chips to China. Nvidia reported $115.8 billion in revenue, up 103% year over year, while its stock has fallen more than 10% since Huang departed.

Why: The Megaspeed thread is the one that touches builders in this region: a Singapore-based data company under a U.S. Commerce investigation since late 2024 for alleged chip diversion to China means GPU capacity sourced through regional brokers and resellers now carries real compliance risk. If you rent GPU capacity in Southeast Asia, ask the provider where the hardware originated and whether they can show end-use/export documentation before you sign — the revenue number (up 103%) says demand and supply are intact, so the risk here is regulatory and reputational, not a shortage.

17 Nov 2026, 9:25 PMCNBC Technology6.0 Nvidia's Colette Kress confronts key questions in first earnings report since Jensen Huang’s exit

Nvidia reports quarterly earnings Tuesday, days after founder Jensen Huang resigned following 33 years at the company; CFO Colette Kress (CFO since 2013) is interim CEO while the board searches for a successor. CNBC flags the questions Kress faces: chip sales to China, whether much of Nvidia's Southeast Asia sales are actually destined for Chinese companies, and risks tied to deals with neoclouds. Nvidia shares have fallen for six straight weeks, the longest losing streak since 2012.

Why: The Southeast Asia question is the one that touches builders here: if regional GPU purchases are being scrutinised as a possible route to Chinese buyers, then GPU capacity bought through SEA cloud/neocloud providers can come with allocation and compliance risk you did not sign up for. Before committing to a multi-month GPU contract, ask the provider where the hardware was allocated from and what happens to your contract if export scrutiny changes — and treat the six-week share slide plus an interim CEO as a sign that supply and pricing terms may shift at the next earnings call.

17 Nov 2026, 10:00 PMTom's Hardware5.5 Nvidia warns customers of new US rule that could expose them to lawsuits for serving Chinese clients

Tom's Hardware reports that Nvidia has warned its customers about new US Commerce Department interim regulations on advanced compute built on American technology, warning that serving Chinese clients could expose those customers to lawsuits. The item as provided contains only the headline and the article's membership/newsletter boilerplate — no rule text, thresholds, effective dates, penalty details, or Nvidia's actual wording.

Why: If you rent, resell, or build on US-origin GPUs or cloud capacity, the headline implies end-use/end-customer obligations may now attach to you rather than only to Nvidia — but this text gives no thresholds or dates, so you cannot act on it yet. The concrete next step is to go read the actual interim rule from Commerce and check whether your own customer contracts and end-use attestations name who your downstream clients are; do not draft policy from this excerpt.

16 Nov 2026, 11:00 PMTom's Hardware5.5 Anthropic says Claude Mythos Preview cut AMD MI355X onboarding from up to 12 months to around three days

Tom's Hardware reports (16 Nov 2026, Shane Downing) that Anthropic claims its Claude Mythos Preview model cut the software work to onboard AMD MI355X GPUs from as much as 12 months down to roughly three days, writing almost 100% of the kernels autonomously. The piece frames this as eroding Nvidia's CUDA moat: if AI can write the low-level code that ties models to a vendor's hardware, the software ecosystem lock-in matters less. The available text is largely subscription boilerplate — no methodology, benchmark numbers, kernel counts, correctness or performance-parity data, and no independent verification.

Why: This is an Anthropic claim about its own model, relayed by a paywalled news-analysis piece, so treat the 12-months-to-3-days figure as unverified until you see the kernel pass rates and performance parity versus the Nvidia path. Practically: if you are budgeting GPU portability work or pricing AMD capacity into a deployment, do not plan against this number yet — ask vendors for correctness and throughput evidence on your own model, not a headline. If it holds even partially, the skill that devalues is hand-writing porting glue, not debugging why a kernel is slow or wrong.

16 Nov 2026, 9:00 PMTom's Hardware5.5 Jailbroken PS4 runs 9-billion-parameter AI model at up to 43 tokens per second

Tom's Hardware reports that a jailbroken PS4 can run a 9-billion-parameter AI model at up to 43 tokens per second using an open-source project called Garlic LLM, which bypasses standard drivers to communicate directly with the console's 2013-era GPU. The supplied page text is almost entirely site navigation and subscription boilerplate, so the model used, quantization, memory footprint, and setup steps are not stated. Treat this as a headline-level report only.

Why: The one usable signal is the ratio: 43 tok/s for a 9B model on a decade-old console GPU implies the software stack — drivers and kernels — is often the binding constraint, not raw silicon. That is relevant if you are sizing cheap or used hardware for local inference instead of buying new GPUs. But with no memory footprint, quantization, or context-length numbers in the text, you cannot yet judge whether the result is reproducible or comparable to a normal llama.cpp setup, so do not plan a build around it.

19 Nov 2026, 5:06 AMCNBC Technology4.5 Nvidia sinks after earnings, dragging down neoclouds CoreWeave, Iren

Nvidia's Q3 revenue came in at $115.84 billion against a $109.10 billion analyst consensus, with Q4 guidance of $130 billion versus $125.1 billion expected, and the stock rose 3.72% in Wednesday premarket trading. Neoclouds moved with it: CoreWeave +6.1%, Nebius +8.2%, Iren +7.2%. The same article notes CEO Jensen Huang resigned the prior week, made his first public statement saying he was open to returning, and that new board chair Harvey Jones said the board does not intend to reinstate him; interim CEO Colette Kress told CNBC that hyperscalers are 'feverishly looking for more focus in terms of the compute that we have available.'

Why: Two concrete signals for anyone renting GPU capacity: Nvidia's Q4 guide of $130 billion implies supply keeps expanding rather than tightening, and a sitting Nvidia CEO transition (Huang out, Kress interim, board refusing reinstatement) is an unusual governance event for the vendor whose roadmap your inference stack depends on. The article's own headline says Nvidia 'sinks' while the body reports a 3.72% premarket gain — treat the sentiment framing as unreliable and check the numbers, not the headline, before reacting.

19 Nov 2026, 5:00 AMTom's Hardware4.5 Nvidia GeForce driver bug reportedly hammers SSDs with up to 30GB of writes per hour

Tom's Hardware reports that users have traced heavy idle SSD wear — up to 30GB of writes per hour — to an Nvidia GeForce driver background service, and that a newer driver is reportedly still affected. The available text is only the headline and page boilerplate: it names no specific driver versions, no Nvidia response, and no confirmed fix or workaround.

Why: If you run a GeForce card in a dev workstation or a local-LLM rig that idles 24/7, 30GB/hour of background writes is a real endurance cost on consumer SSDs — but with no version numbers in this report you can't tell whether you're affected. Measure before acting: check your drive's Total Host Writes / SMART data over a day, then again a day later, instead of switching drivers on an unconfirmed report.

16 Nov 2026, 7:00 PMTom's Hardware4.5 Nvidia amassed $530 billion in financial guarantees to corner the AI market

Tom's Hardware's headline states Nvidia has accumulated $530 billion in financial guarantees, with supply commitments rising by nearly a quarter of a trillion dollars (~$250B) in a single quarter. The supplied page text, however, is almost entirely site navigation and subscription boilerplate — it contains no reporting on how the guarantees are structured, who they are with, or what terms apply.

Why: The only usable facts here are the two headline figures: $530B in total guarantees and a ~$250B single-quarter jump in supply commitments. That is a signal about how much AI compute capacity is being locked in years ahead, which is relevant if you are budgeting GPU/cloud spend or modelling whether inference prices keep falling — but the text gives no breakdown, so do not treat it as a basis for any decision yet. Wait for the underlying filing or full article before drawing conclusions.

17 Nov 2026, 9:30 PMTom's Hardware4.0 US government confirms criminal investigation into Singapore company at the heart of Jensen Huang scandal

Tom's Hardware reports that the US government has confirmed a criminal investigation into Megaspeed, a Singapore-based company, and that Nvidia has placed the firm on a 'no sell' list over 'suspicious activity.' The article is framed as being at the centre of a 'Jensen Huang scandal,' but the supplied text contains only site navigation and subscription boilerplate — no agency name, no charges, no dates, no product lines, no volumes, and no statement from Nvidia, Megaspeed, or the US government.

Why: There is not enough in this text to act on. What is usable is narrow: a Singapore-domiciled entity appears to have been cut off from buying Nvidia hardware, which is the kind of counterparty risk that matters if you or a portfolio company sources GPU capacity through a Singapore intermediary. Before signing or renewing any brokered GPU/cloud capacity contract, ask the supplier directly which Nvidia SKUs they can lawfully resell and under what US export terms — but treat the Megaspeed story itself as unverified detail until the primary filing or a fuller report is available.

17 Nov 2026, 5:50 AMTom's Hardware4.0 New US rules aim to restrict Chinese firms’ remote access to AI servers and infrastructure

Tom's Hardware reports that new US rules are intended to close a loophole and restrict Chinese firms' remote access to AI servers and infrastructure, specifically targeting cloud-rental workarounds to existing GPU export controls. The fetched article text contains only site navigation and subscription prompts, so no details on the issuing agency, effective date, compute thresholds, or enforcement mechanism are available here.

Why: There is not enough in this text to tell you what to change: no effective date, no definition of 'remote access', no threshold, no named agency. The only defensible takeaway is procedural — if you rent GPU capacity through resellers or intermediaries whose ownership or jurisdiction you have not verified, that due diligence now sits on a stated US policy intent rather than a hypothetical. Treat the headline as a signal to read the actual rule text before signing a multi-year capacity commitment, not as a reason to migrate workloads today.

17 Nov 2026, 5:00 AMTom's Hardware4.0 AMD grabs over 40% of new TOP500 list for supercomputers, up 14% since last year

Tom's Hardware reports that AMD now accounts for more than 40% of the new TOP500 supercomputer list, a 14% increase since last year. The new Lux system, built with the Department of Energy, debuts at No. 14 using AMD EPYC CPUs and Instinct MI355X GPUs, and is the first of two machines under a $1 billion AMD–DOE deal. The article gives no benchmark figures, pricing, or availability details beyond those counts.

Why: The only actionable signal here is supply diversification: AMD's MI355X showing up in a top-14 DOE machine, backed by a $1B two-system deal, is evidence that non-Nvidia accelerator capacity will keep expanding. But the text contains zero performance, price, or availability data, so nobody should change procurement or training plans off this headline alone — treat it as a marker to re-check GPU pricing when MI355X cloud instances actually appear.

18 Nov 2026, 7:00 PMTom's Hardware3.0 China threatens 'countermeasures' in the face of new AI GPU rental and export restrictions

Tom's Hardware reports that Beijing is threatening "countermeasures" in response to a new round of AI GPU rental and export restrictions, with retaliation talk timed ahead of a Thursday meeting involving Xi Jinping. The excerpt itself is almost entirely subscription/navigation boilerplate: it does not name the specific rule, the agency issuing it, the GPUs covered, the effective date, or the rental mechanism being targeted.

Why: There is not enough here to act on. The text names no GPU models, no performance thresholds, no rental-provider categories, and no implementation date, so a team sourcing cloud GPU capacity cannot tell whether its current provider or instance type is affected. The only decision supported right now is procedural: wait for the actual rule text (scope, thresholds, effective date) before repricing or re-architecting anything.

17 Nov 2026, 10:30 PMTom's Hardware3.0 AMD attacks Nvidia's FP64 emulation at SC26, touts native double-precision on Instinct MI430X

At SC26, AMD criticized Nvidia's FP64 emulation and promoted native double-precision on its Instinct MI430X. AMD claimed Ozaki-style emulation fits less than 10% of HPC workloads, roughly triples memory use, and is a poor fit for many HPC workloads.

Why: If you are sizing GPU nodes for FP64-dependent HPC or scientific code, AMD's stated under-10% fit rate and 3x memory penalty are the concrete claims to test against your own kernels; do not treat emulated FP64 as a drop-in substitute without workload-specific benchmarks.

17 Nov 2026, 3:05 AMTom's Hardware3.0 Nvidia announces Rubin HPC GPU with 'record number of CUDA cores' at Supercomputing 2026

At Supercomputing 2026, Nvidia announced a Rubin HPC GPU that it says packs "over 50,000" CUDA cores on a single chip, which Tom's Hardware reads as implying a four-die layout. Nvidia is pitching it at "agentic HPC workloads," but the chip does not arrive until 2028. The piece is a news-analysis with no pricing, benchmarks, memory specs, or named customers.

Why: There is nothing here to act on yet: no price, no benchmarks, no memory or interconnect details, and a 2028 ship date. If you are sizing GPU capacity, cloud commitments, or inference budgets for the next 12-24 months, plan against hardware you can actually rent or buy now rather than waiting on Rubin HPC. Treat "agentic HPC" as an unquantified marketing label until Nvidia publishes numbers.

16 Nov 2026, 7:15 PMTom's Hardware2.5 Grab a DDR5 gaming PC with a 10GB GPU for only $949

Tom's Hardware flags a deal on iBuypower's Slate Gaming desktop at $949, pairing an Intel B570 graphics card with 10GB of VRAM and 16GB of DDR5 memory. The item is a retail price note rather than a review or benchmark, and the supplied text is mostly site navigation and membership boilerplate — no performance numbers, no test results, no availability date beyond the November 16, 2026 publish stamp.

Why: For anyone considering a cheap local box for small-model inference or Stable Diffusion, 10GB of VRAM at $949 is the only decision-relevant datapoint here — and it is unverified, since the text carries no benchmarks, no power figures, and no confirmation of Malaysian pricing or availability. Treat it as a price bookmark, not a purchase recommendation; import duty, shipping, and warranty support would all need checking before this beats a local listing.

16 Nov 2026, 5:32 PMThe Hacker News2.5 SANSAR Linux Cryptomining Malware Uses Modified NVIDIA Library to Hide GPU Activity

The Hacker News headline reports a Linux cryptomining malware family called SANSAR that reportedly modifies an NVIDIA library so GPU activity stays hidden. The excerpt supplied here is only page navigation and the headline itself — no affected versions, indicators of compromise, mining pool details, loader mechanism, or vendor response are present. Treat this as a headline, not a briefing: nothing in the text lets you verify or act on the claim.

Why: There is no actionable detail in this text, so don't change anything based on it yet. The one implication the headline does support: if the malware patches the NVIDIA library itself, then the usual 'run nvidia-smi and look at utilization' check on a GPU box may not show the load — meaning unexplained cloud GPU spend or power draw, not GPU metrics, would be your only signal. Wait for the full technical write-up before deciding whether this touches your fleet.

18 Nov 2026, 8:16 AMTechCrunch2.0 Nvidia promises that revenue will keep soaring — for now

Nvidia's third quarter of fiscal 2027 came in at $115.8 billion in revenue, up 103% year over year and above its own $108 billion guidance, per interim CEO Colette Kress on the Tuesday call. Kress guided the fourth quarter to $130 billion (±3%) and raised the fiscal 2028 outlook to roughly 75% year-over-year growth, up from 70%. The report follows the departure of founder and CEO Jensen Huang the previous week, with the stock down about 10% since then and up about 4% after hours; Nvidia also spent $28.1 billion on buybacks in the quarter, leaving $227.3 billion on a program increased by $150 billion in September 2026.

Why: The only builder-relevant number here is the demand signal: Nvidia guiding to a $130 billion quarter and lifting FY2028 growth to ~75% implies it does not expect GPU supply or inference pricing to loosen soon, so any capacity or per-token cost assumption you baked into a 2027 budget is probably too optimistic. Beyond that budgeting input, the article contains no product, pricing, API, or availability changes — there is nothing here to change your stack over.

17 Nov 2026, 7:14 PMTom's Hardware2.0 Save $385 on this 4K-ready Alienware gaming PC with an RTX 5080 and 32GB DDR5 right now

Tom's Hardware flags an Alienware Aurora ACT1250 gaming desktop at $2,799.99, a $385 discount. The configuration includes an RTX 5080, a 24-core Intel CPU, 32GB DDR5, and a 1TB SSD, and is described as 4K-ready.

Why: This is only a US retail deal reference: there is no Malaysian pricing, warranty, shipping, or local availability in the text, so builders should not treat $2,799.99 as the real cost of a local GPU workstation. It may be useful only for comparing the raw spec bundle against local quotes.

Top