Big AI Went Vertical.

Every AI headline this week was really about who owns the pipes.

“Anthropic confirmed on 5 August that it is standing up an in-house custom silicon team led by Clive Chan — previously the second hardware engineer on OpenAI's own chip programme — targeting roughly 50% cuts in per-token inference cost, with Samsung reportedly scouted as a manufacturing partner. OpenAI's public S-1 prospectus is expected between now and late August ahead of a September IPO target above US$1 trillion, up from an US$852 billion private valuation. Alphabet raised its 2026 capital-expenditure guidance to US$195–205 billion after Google Cloud's Q2 revenue jumped 82% to US$24.8 billion and cloud backlog hit US$514 billion. At home, LHDN's MyInvois engine started enforcing the new 12-character SSM BRN format on 1 August with no grace period, and Meta's WhatsApp Business Agent pricing — roughly four to five sen per message — went live for Malaysian WhatsApp Business accounts the same day.”

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

The Big Three

1. Anthropic Confirmed It Is Building Its Own Chip

On 5 August, Anthropic confirmed to reporters what its job listings had already implied: the company is standing up an in-house custom silicon team, offering roles at US$320,000–485,000 to work on chips and Claude models jointly. Clive Chan — Anthropic's June 2026 hire, previously the second hardware engineer on OpenAI's own chip programme after a stint at Tesla's Dojo effort — is leading it. The Information reported that Samsung is one manufacturing partner being scouted. Anthropic said it will keep buying from Nvidia, AMD, AWS Trainium and Google TPU; the custom silicon is meant to add a layer, not replace what already works. The stated target is to roughly halve the cost of running Claude per token.

So what for a business owner? In the last six months every major AI vendor has moved down the stack — OpenAI, Google, Amazon and now Anthropic all have chip programmes in some form. The upside is real: cheaper inference in 2027 flows through to cheaper AI features in the tools you already pay for. The downside is that the same vendors will bundle harder, discount their own end-to-end stack, and make it more painful to switch off it later. If you are looking at AI-heavy software right now, avoid three- and five-year lock-ins where you can. The cost curve is moving fast, and today's list price will look expensive by this time next year.

2. OpenAI's IPO Paperwork Is About to Go Public

OpenAI filed a confidential draft S-1 with the SEC on 8 June. That gets the review going; it does not make the numbers public. The full prospectus — revenue, unit economics, cost of goods sold, whether the business actually makes money at scale — is expected between now and late August, with a September listing target and a mooted valuation above US$1 trillion, up from the US$852 billion private round closed earlier this year. Goldman Sachs and Morgan Stanley are on the deal. No listing date has been announced, and "confidential filing" is not the same as an effective registration statement — but the window is close.

So what for a business owner? Two things change once OpenAI is a listed company. First, the numbers you and your accountant have never seen — how much a subscription actually costs OpenAI to run, how the free tier is subsidised, whether ChatGPT Enterprise carries gross margin — become public and quotable. That is useful ammunition when you are negotiating with any AI vendor: everyone's costs will suddenly have a public benchmark. Second, listed companies have to show quarterly growth. Expect OpenAI to lean harder into paid features and enterprise contracts and further away from generous free tiers. If any part of your business quietly runs on ChatGPT Free, budget for that to become ChatGPT Something-You-Pay-For inside the next twelve months.

3. Alphabet Now Plans To Spend US$205 Billion This Year

At Q2 earnings on 22 July, and reaffirmed in early-August briefings, Alphabet lifted 2026 capital-expenditure guidance to US$195–205 billion, up from US$180–190 billion. Google Cloud revenue grew 82% year on year to US$24.8 billion in the quarter. Cloud backlog — signed contracts not yet delivered — reached US$514 billion, of which management expects at least half to convert within 24 months. CFO Anat Ashkenazi's phrase was blunt: "demand still outpaces that investment." Amazon is projected to spend roughly the same figure in 2026; Microsoft and Meta together push the hyperscaler total past US$700 billion for the year.

So what for a business owner? Two useful signals. The scale of committed enterprise cloud spending tells you the direction is not reversing — companies larger and better-informed than yours have pre-paid for years of AI compute. That is not a reason to buy anything, but it is a reason to stop waiting for AI to "blow over" before you engage with it. The second signal is quieter: at this spend, hyperscalers can subsidise. Google, Microsoft and AWS will keep bundling generous AI features into their small-business tiers of Workspace, 365 and AWS in the coming quarters to pull SMEs into their ecosystem. Those bundles — not standalone AI vendors — are the practical entry point for most Malaysian SMEs: cheaper, better-supported, and already inside tools your team uses.

Closer to Home: Malaysia

The single most consequential Malaysian development this week is not a headline story — it is a MyInvois validation change that took effect on 1 August. LHDN's MyInvois engine now validates the Business Registration Number (BRN) on every submission against the 12-character SSM standard (e.g. 202500001234). Any invoice submitted with a non-conforming BRN — an older 6-, 7-, 9- or 10-character code, or a mismatch against HASiL's records — is rejected outright. There is no grace period on this specific rule. Sellers must obtain and verify the correct BRN from every buyer before issuing an e-invoice; a mismatch is a stalled invoice, and a stalled invoice is a delayed payment.

That sits inside the broader e-invoice picture, which has not changed this week and is worth repeating honestly: the Phase 4 (RM1–5 million) grace period on penalties runs to 31 December 2027, with the RM200–20,000 per-invoice penalty regime kicking in from 1 January 2028. But transactions of RM10,000 or more already require an individual e-invoice today, and LHDN's own February reporting caught more than 500,000 non-compliant cases and RM1.4 billion in unreported income. The grace period covers penalties, not audits.

The other Malaysian item worth flagging: Meta's WhatsApp Business Agent pricing went live the same day for Malaysian accounts, billed directly in ringgit. Reported rates translate to roughly four to five sen per exchanged message depending on complexity. From 1 October 2026, Meta will also start charging for service messages that were previously free. If WhatsApp is your primary sales channel — and for most retail, F&B and service SMEs in the RM1–20 million range it is — this is now a line item worth putting on a spreadsheet before it becomes a surprise on your first monthly invoice.

Nothing dated from Singapore this week rises to the "materially affects a Malaysian SME" bar, so we will leave it alone this round.

What This Means for Your Business

1. Verify your top-40 counterparties' BRNs this month. If you are subject to e-invoicing, pull your top 20 customers and top 20 suppliers, get the 12-character SSM BRN from each in writing, and update your accounting system. This is now the single most likely reason your next e-invoice will fail to submit — and a failed submission on a RM10,000+ transaction is an audit-flagged missing document. Two hours of admin this week saves a fortnight of chasing later.
2. Budget for WhatsApp as a cost line, not a free channel. Whatever you do on WhatsApp Business today probably costs you nothing. From 1 August, most of it starts to cost sen per message; from 1 October, service messages join the meter. Estimate your current monthly message volume, multiply by RM0.05–0.10 per message, and put the figure into your P&L before Meta's first ringgit-billed invoice lands. If the number is meaningful, tighten your automations now — templates you actually need, not templates padding your engagement metrics.
3. Do not sign a three-year AI-tools deal. Inference cost has been roughly halving each year at the lower tiers, and Anthropic's chip play accelerates that. A twelve-month renewal on any AI-heavy software gives you a chance to reprice at 2027 rates. A three-year lock-in does not. If a vendor is offering a big discount for a three-year commit right now, that is the vendor telling you it expects to be less competitive later.

The Practical Question

If OpenAI is about to publish what it actually costs to run one ChatGPT subscription, what does that tell me about the price I am being charged for AI features in the tools I already own?

You will not have to interpret the S-1 yourself. The financial press will do it inside twenty-four hours of the filing. Set a simple alert this week — "OpenAI S-1" on Google News, one email a day — and wait. When the number lands, look up what you personally pay per user per month for any AI-embedded tool: your CRM, your accounting suite, your customer-support platform, your WhatsApp automation. If the ratio between what you pay and what OpenAI's own cost per subscription turns out to be is more than about five to one, your vendor is charging a premium that will not survive the next two renewal cycles. Better to have that data point in your pocket before the next contract conversation than after it.

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 →