The Cloud Got a Marketplace. LHDN Got Analytics.31 posts
← Back to InsightsAI Trends

The Cloud Got a Marketplace. LHDN Got Analytics.

October 23, 2026·7 min read

Three major AI stories broke on the same Tuesday, and LHDN quietly confirmed the fourth.

On 21 October, OpenAI's flagship models — GPT-Astra, GPT-6.1 Sol and GPT-6 Luna, plus Codex — went live in Google Cloud's Model Garden, a third major cloud for buying them. The same day, Anthropic opened a waitlist for Claude Duet, a proactive agent on Opus 5.5 that reviews your email every morning and drafts replies for approval. On 22 October, American Express announced about 3,500 job cuts citing AI preparation, and OpenAI's bridge round closed at a US$1.25 trillion valuation — US$150 billion below its pitch. In Putrajaya, LHDN confirmed 2.09 billion e-invoices from 269,000+ taxpayers and that descriptive, diagnostic, predictive and prescriptive analytics now run on every submission.

Here's what happened, why it matters, and what your business should do about it.


The Big Three

1. OpenAI's models landed on Google Cloud

On 21 October, OpenAI announced that its flagship models — GPT-Astra, GPT-6.1 Sol and GPT-6 Luna, plus Codex — will be sold through Google Cloud's Model Garden, in private preview now with broader availability "in the coming weeks." Microsoft remains the primary cloud partner, but this is a third commercial channel after Azure and AWS.

The commercial detail that matters most: Google Cloud customers can apply existing spend commitments to OpenAI purchases. If your finance team already has Google Cloud credits, those credits now buy GPT. OpenAI sits beside Gemini and Claude in the same marketplace — the first time enterprise buyers can benchmark the three major frontier families on one contract and switch between them without moving workloads.

So what for a business owner: Vendor lock-in just lost most of its grip at the top of the stack. The cloud is no longer the deciding factor between model providers — the task is. A Malaysian SME on Google Workspace that was quietly putting off GPT access because switching clouds felt heavy no longer has the excuse.

2. Anthropic shipped a proactive AI agent

Also on 21 October, Anthropic opened a waitlist for Claude Duet, a research preview of a proactive agent on Opus 5.5. Unlike the chat window, Duet runs on its own cloud computer, reviews your email and calendar every morning, and sends a brief on what needs attention. It drafts replies for review, chases unpaid Stripe invoices, and handles longer jobs — research briefs, slide decks — in the background. Before it sends a message, makes a payment, signs a document, or deletes anything, it asks for approval.

It is included in Pro and Max plans at no extra cost, by invitation to US subscribers for now. Meta's Muse and OpenAI's Dots are the direct competitors — the vendor race has shifted from "which model answers best" to "which agent runs quietly in the background."

So what for a business owner: The baseline for an AI assistant just shifted from reactive to proactive. If you employ anyone whose job is checking email in the morning and surfacing what matters, that job is now partly done by software. The question is whether you set the approval rules — which inboxes, which systems, which actions need a human — or let someone else's defaults do it.

3. The AI payroll cuts started attaching a headcount

On 22 October, American Express announced it would cut about 3,500 jobs — roughly 5% of its workforce — citing AI preparation. Barclays made a similar announcement the day before. The same morning, CNBC reported OpenAI's bridge round is closing at a US$1.25 trillion pre-money valuation — US$150 billion below the US$1.4 trillion it had pitched earlier.

Two things to hold at once. The job cuts are real — these are the first large banks to attach a headcount number to the AI reason. The valuation gap is also real — even OpenAI has a ceiling buyers won't pay through.

So what for a business owner: If a US bank is cutting 3,500 roles on an AI justification, your competitors — the ones who sell to you and the ones who buy from you — are reading the same memo. Expect their cost structures to shift through 2027, and pricing with them. When trillion-dollar labs miss their own fundraising targets, treat vendor pricing today as negotiable, not fixed.


Closer to Home: Malaysia

LHDN announced on 21 October that it has now received more than 2.09 billion e-invoices from over 269,000 taxpayers since the national rollout. More importantly, it confirmed what every finance team suspected since January: the analytics module is live and running on every submission.

The system runs four kinds of analysis — descriptive (what happened), diagnostic (why), predictive (what's likely next), and prescriptive (what to do about it) — and reviews each taxpayer's financial profile when invoices come in. LHDN added that a flagged pattern does not mean fault; variations "may come from account adjustments or reporting timelines." The practical implication is still clear: anomalies now surface by default. You don't need an auditor to find them; the system does it first.

LHDN also named the common non-compliance patterns it is finding. Businesses issue e-invoices for stock sales but miss other taxable activities — disposing of business assets, earning rental income, and getting the self-billed e-invoice flow wrong when paying suppliers still exempt. Not exotic errors. The kind that happens quietly when the finance team is busy and nobody wrote down the rule.

The counterweight: the e-Invoice Special Voluntary Disclosure Programme (SVDP) runs until 31 December 2027 and offers a full penalty waiver for miscalculations and clerical errors, provided the correction is transparent and not deliberate manipulation. The RM3 million exemption for MSMEs stays. The RM10,000-per-transaction rule — every business, exempt or not — is still live. Phase 4 (RM1–5M turnover) penalties of RM200 to RM20,000 per invoice start on 1 January 2028.

Singapore note: No regionally material Singapore SME story warranted inclusion this week.


What This Means for Your Business

  1. Use the voluntary disclosure window before the analytics engine does. LHDN has publicly named the errors it catches most often: unreported asset disposals, missed rental income, self-billing confusion. If any ring a bell in your past 12 months, the SVDP window closes 31 December 2027 with a full penalty waiver. If the system flags first, you lose that waiver — and from 1 January 2028, penalties run RM200 to RM20,000 per invoice. Walk your bookkeeper through asset disposals, rental income and self-billing records this month, not next quarter.
  2. Decide who gets to click "send" before you install an agent. The baseline AI assistant now watches email, drafts replies, and chases invoices. Before you let one in, write down which actions need a human nod: drafting is fine, sending is approval, payment is always approval. Anthropic's own product keeps those checkpoints by default. Make sure whatever you deploy keeps them too — including the quiet ones a setup wizard might helpfully turn off.
  3. Procure models by task, not by vendor. With GPT, Claude and Gemini all on every major cloud, the question is no longer "which cloud do we standardise on" but "which model fits which job." Proactive email and invoice agent: Claude. Google Workspace workflow: Gemini. Codex and GPT-6 engineering, now via Google Cloud: OpenAI. The RM30M MDEC grant from Budget 2027 covers whichever one makes sense — but only if you can explain the job it is doing.

The Practical Question

If LHDN's analytics engine ran a full pass on your last 12 months of invoices this afternoon, what would it find that you'd rather have fixed yourself?

Most owners will instinctively answer "nothing" and move on. The ones worth watching this quarter will take ten minutes with their accountant and actually check — asset disposals, rental income, self-billing, voids, credit notes, late submissions. The SVDP exists because LHDN would rather you fix the errors than it does. From 1 January 2028, that preference reverses.


At The Empyrean, we help Malaysian SMEs find the practical, repeatable tasks where AI delivers value without disruption. If you're not sure where to start, we're happy to take a look at your operations and tell you honestly what would make sense.

Talk to us →