Ternus era began at Apple on 1 September 2026, and it began in the same week that Nvidia agreed to spend $12.93 billion buying the world’s largest open model hub. Two of the most valuable companies on earth answered the same question — how do you win artificial intelligence? — and gave opposite answers within days of each other. Apple decided to rent its intelligence and keep its distribution. Nvidia decided to buy the entire stack it sells into.

John Ternus took over from Tim Cook after a 15-year run, and Cook moved to executive chairman. Three days later Jensen Huang announced the Hugging Face deal, the largest acquisition in Nvidia’s history. Neither event was a surprise on its own. Put side by side, the opening of the Ternus era and Nvidia’s buying spree mark the moment the industry stopped pretending there is one obvious way to build an AI business.

Both bets reduce to the same component: a large language model good enough that customers stay, and a bill for running it that somebody has to absorb. Apple has chosen to rent that component. Nvidia has chosen to own the ground beneath it.

This article sets out what each company actually committed to, the money behind both bets, where each one could break, and what any of it changes for an organisation deciding how much of its own AI stack to own. We track releases like these in our AI models and tools hub as they land. The short version: the Ternus era is buying optionality at roughly $1 billion a year, Nvidia is buying certainty at roughly forty times that, and only one of them can be wrong cheaply.

What the Ternus Era Actually Inherits

apple ternus era nvidia bets whole ai stack b two pan balance scale beam and dishes

The handover was announced in April 2026 and executed on 1 September. It was not a crisis succession, and the Ternus era does not begin with a company in trouble.

The handover on paper

Apple’s board approved the transition unanimously after what the company described as a long-term succession process. Cook stayed through the summer working alongside his successor, then became executive chairman with a remit that includes engaging policymakers. Arthur Levinson, non-executive chairman for 15 years, became lead independent director on the same day.

That is an unusually orderly start. The Ternus era inherits a balance sheet, a board and a product calendar that were all set before the new chief executive had the title. Very few Ternus era decisions of consequence can be made before the September event has already happened.

A 25-year hardware career

Ternus is 51 and joined Apple’s product design team in 2001. He was appointed vice president of hardware engineering in 2013 and promoted to senior vice president in 2021. Over that span he oversaw the launch of the iPad, AirPods, Apple Watch and Vision Pro, plus the Apple Silicon transition away from Intel.

That last item matters more than the rest for what comes next. Moving the Mac to Apple’s own chips was a multi-year vertical integration project executed on schedule, and it is the single best evidence anyone has about how the Ternus era might approach the parts of AI that Apple wants to own.

The first test is eight days away

Cook’s successor sent a short memo to staff on his first day. He thanked Cook for leading “with his values and with such decency and humanity,” then flagged the immediate job: “We have a huge launch next week that’s going to be phenomenal.”

That launch is the “Surprise and Shine” event on 9 September 2026 at the Steve Jobs Theater. It is expected to carry the first foldable iPhone, the iPhone 18 Pro line, Apple Watch Series 12 and Ultra 4, and iOS 27. It is also the first public verdict on the Ternus era, delivered eight days into it.

No chief executive gets to choose their first launch, and the Ternus era did not choose this one. What it does get to choose is everything that follows.

The Ternus Era Bet: Rent the Intelligence, Own the Devices

apple ternus era nvidia bets whole ai stack c four tier stacked block tower

Apple’s AI strategy was set before the handover, which is the awkward part. The Ternus era inherits a bet it did not place.

The Gemini deal in numbers

Apple and Google announced a multi-year partnership on 12 January 2026 under which Google’s Gemini models become the foundation of a rebuilt Siri. Apple is reported to be paying roughly $1 billion a year for a custom 1.2-trillion-parameter model. Requests route through Apple’s Private Cloud Compute layer — Apple Silicon servers acting as a buffer — with Apple stating that interactions are anonymised and never used to train Google’s models.

Apple has not stopped building. It continues to develop its own cloud model in the region of a trillion parameters, which could in principle power Siri independently. The licence buys time, not a permanent answer, and how much time is a Ternus era judgement call.

What “capex-light” actually means

The financial shape of this choice is stark. In the nine months to the June 2026 quarter, Apple spent about $6.8 billion in capital expenditure, an annualised run rate near $9 billion. In the single June 2026 quarter, Amazon spent $54.2 billion, Alphabet $44.9 billion and Meta $30.1 billion.

Apple’s nine-month infrastructure bill is smaller than what any one of those three spends in a quarter. That is the Ternus era inheritance in one comparison: a company monetising AI at the end of the stack rather than building it.

Capital expenditure: three hyperscalers in one quarter versus Apple in nine months
Amazon, Q2 2026 $54.2bn
Alphabet, Q2 2026 $44.9bn
Meta, Q2 2026 $30.1bn
Apple, nine months to June 2026 $6.8bn

The business it protects

The strategy is not timidity. Apple’s June 2026 quarter produced $109.4 billion of revenue, up 16%, with a 50.1% gross margin and $29.8 billion of net income. iPhone revenue rose 22% to $54.3 billion, Mac 29% to $10.4 billion, and Services 12% to a record $30.7 billion. Research and development hit a record $11.73 billion, up 32%.

Read those together and the logic is clear. Apple spends heavily on the thing it sells and lightly on the thing it buys. The Ternus era starts with the highest-margin distribution network in consumer technology and a supplier relationship where the intelligence comes from.

Whether that is discipline or complacency is the argument the Ternus era will be judged on, and the evidence will not arrive for several years.

MeasureApple, quarter to June 2026Nvidia, quarter to July 2026
Revenue$109.4bn, up 16%$96.2bn, up 106%
Largest segmentiPhone, $54.3bnData centre, $89.0bn
Gross margin50.1%75.0%
Diluted EPS$2.02$2.46 GAAP
AI postureLicence the model, own the deviceOwn the silicon, buy the layers above
Annual AI outlay in view~$1bn Gemini licenceTens of billions in equity and M&A

Nvidia's Opposite Bet: Buy the Whole AI Stack

apple ternus era nvidia bets whole ai stack d desk bell dome and round base

While Apple was changing chief executive, Nvidia was finishing a year of buying everything above its own chips.

The Hugging Face acquisition

On 3 September 2026 Nvidia agreed to acquire Hugging Face for $12.93 billion — $11.9 billion to shareholders and $1 billion in equity to retain staff joining Nvidia, with completion expected in the first half of next year. The platform brings more than 18 million developers, over 3 million models, more than 500,000 datasets, over 1 million applications and around 200,000 companies.

Nvidia was already the largest single contributor there, with 500-plus open models and 250-plus open datasets published. We covered the terms, the valuation multiple and the antitrust exposure in detail in our analysis of the Nvidia Hugging Face acquisition.

The rest of the 2026 shopping list

Hugging Face was the largest item, not the only one. Across 2026 Nvidia committed roughly $30 billion to OpenAI and up to $10 billion to Anthropic in February, backed xAI in January, and was reported in September to be in talks for about $2.5 billion in Mira Murati’s Thinking Machines Lab. It put $2 billion each into CoreWeave, Nebius, Marvell, Lumentum and Coherent, up to $2.1 billion in IREN warrants, around $3.2 billion into Corning, $3.5 billion into MediaTek convertible bonds, and $6 billion in licensing plus $1 billion of equity into Poolside.

Bank of America put Nvidia’s equity commitments to AI-lab partners at as much as $70 billion early in 2026. Huang said in March that the OpenAI and Anthropic stakes would be the final chapter of direct equity investment in the major labs. The Hugging Face deal is a different instrument entirely — an outright purchase of a layer, not a stake in a customer.

TargetLayerAmountAnnounced
OpenAIModels$30bnFeb 2026
AnthropicModelsUp to $10bnFeb 2026
Thinking Machines LabModels~$2.5bn, in talksSep 2026
Hugging FaceSoftware and distribution$12.93bnSep 2026
PoolsideSoftware$6bn licence + $1bn equityAug 2026
MediaTekChips$3.5bn convertiblesAug 2026
CorningOptics~$3.2bnMay 2026
CoreWeave, Nebius, IRENCloud$2bn, $2bn, up to $2.1bnJan–May 2026

Where the money actually went

Group those figures by layer and the shape of the strategy appears. Models take $30bn plus $10bn plus $2.5bn, or $42.5 billion. Software and distribution take $12.93bn plus $6bn plus $1bn, or $19.9 billion. Chips and optics take $3.2bn plus $2bn plus $2bn plus $2bn plus $3.5bn, or $12.7 billion. Cloud capacity takes $2bn plus $2bn plus $2.1bn, or $6.1 billion.

Nvidia’s 2026 commitments by layer, summed from the figures above
Model developers $42.5bn
Software and distribution $19.9bn
Chips and optics $12.7bn
Cloud capacity $6.1bn

The quarter that pays for it

Nvidia can afford this because of what the July 2026 quarter produced: $96.2 billion of revenue, up 106% year on year, with data centre revenue of $89.0 billion, up 117%. Gross margin was 75.0% and GAAP diluted earnings were $2.46 a share. Guidance for the following quarter is $108.0 billion, plus or minus 2%.

Huang’s framing on that call is the clearest statement of the thesis: “AI has reached its inflection point. It’s doing useful work. Its tokens are productive and profitable. Now, compute is revenue.” A company that believes compute is revenue will buy every layer that decides where compute gets spent.

Why the Ternus Era and Nvidia's Stack Bet Landed in the Same Week

apple ternus era nvidia bets whole ai stack e hourglass two bulbs in a frame

The timing is coincidence. The symmetry is not.

Two directions through the same diagram

Nvidia is moving up the stack, from silicon into software, model hubs and developer distribution. Model labs are moving down it, into custom accelerators and their own data centres. Apple is doing neither: it sits at the consumer end and buys the intelligence wholesale.

Each is a coherent answer to vertical integration as a strategic question, and they are mutually exclusive. You cannot simultaneously own every layer and stay deliberately capex-light.

What each company is really protecting

Nvidia is protecting the position of its instruction set and toolchain against a future where customers standardise on someone else’s. Buying the place developers publish and download models is a defence of demand, not an expansion of supply.

Apple is protecting margin and installed base. More than two billion active devices is a distribution asset that no model licence can dilute, provided the licensed model is good enough that nobody leaves. That proviso is the entire risk of the strategy the Ternus era has taken on.

Nvidia’s version of the same worry runs the other way. Its installed base is developers rather than consumers, and the Hugging Face purchase is the equivalent of the device moat the Ternus era already has for free.

What the Ternus Era Has to Prove on 9 September

apple ternus era nvidia bets whole ai stack f kettlebell round body and arch handle

Eight days is not much of a honeymoon. The Ternus era gets judged first on a product that was finished before it started.

The Siri problem, restated

Apple promised a rebuilt Siri, missed, promised again, and then bought the capability. Siri is a natural language processing product before it is anything else, and the version Apple shipped for years was visibly behind what the current generation of assistants can do. The 9 September event is where the licensed replacement has to appear and work. Gene Munster of Deepwater Asset Management set the bar bluntly: the new Siri needs to be “10 out of 10” on arrival.

That is a harsher standard than a normal product launch because the delay itself became the story. A merely competent assistant confirms the criticism; only a clearly excellent one retires it. The Ternus era does not get graded on effort here.

The foldable and the hardware case

The other half of the event plays to the new chief executive’s background. A first foldable iPhone is a hardware engineering problem of exactly the kind Ternus spent 25 years solving, and it is the sort of launch where his track record with iPad, AirPods and Apple Silicon is the relevant evidence.

If the Ternus era has an early advantage, this is it. Hardware credibility is the one thing the appointment guarantees.

What to watchWhy it mattersWhen
Rebuilt Siri qualityDecides whether the licence bought parity or just time9 Sep 2026
Foldable iPhone executionTests the hardware case for the appointment9 Sep 2026
Compute cost disclosureCook admitted there is no complete plan for itNext earnings call
Apple’s own trillion-parameter modelDetermines whether the Google dependency is temporary2027 at the earliest
Hugging Face regulatory reviewA block would strand Nvidia’s distribution betFirst half of 2027

Where Nvidia's Whole-Stack Bet Could Break

Buying every layer creates exposure that selling one layer does not.

Antitrust and the neutrality promise

Nvidia’s announcement committed to keeping Hugging Face open to all developers regardless of compute platform, with continued multi-cloud and multi-accelerator support. That promise is necessary precisely because the concern is obvious: the dominant accelerator vendor now owns the neutral ground where models are published.

Regulators do not have to prove intent to act on structure. A deal expected to close in the first half of 2027 has a long window in which that argument can be made — a risk the Ternus era, buying nothing, does not carry at all.

Circular financing

The other exposure is arithmetic. Nvidia invests in labs and clouds, which buy Nvidia hardware, which books revenue that funds further investment. Each step is legitimate on its own terms and the aggregate still looks like a loop. If model demand slows, the loop reverses faster than a hardware-only business would.

Anyone running their own data centre operations has a direct interest in how that unwinds, because it sets accelerator pricing and availability for everyone downstream.

Where the Ternus Era Rental Bet Could Break

Apple’s exposure is narrower but sharper, and it has a name.

The compute bill nobody has costed

On his final earnings call Cook conceded the company does not have a complete plan for what Siri’s compute costs become as usage scales. His own words: “We’ll see what Siri AI does from the cost side of it. There’s also the ability when people use it a lot for them to move up on an iCloud Plan as well, and so what the balance of that is, is a bit uncertain at the moment.”

That is an honest answer and an unresolved one. A $1 billion annual licence is trivial against $109.4 billion of quarterly revenue. A variable compute bill that scales with two billion devices is a different instrument, and pricing it is now the Ternus era’s problem.

Cook also floated the mitigation: heavier Siri users moving up an iCloud tier. That converts an unbounded cost into a subscription question, which is a genuinely Ternus era shaped answer — monetise the device relationship rather than the model.

Dependency on a competitor

The second risk is structural. The intelligence in Apple’s flagship consumer feature comes from a company that also sells phones, an operating system and its own assistant. Apple has managed that shape before with search, profitably, for two decades — but search was a default placement Apple was paid for. This is a capability Apple pays for, and reversing it means shipping a model of its own.

The catch-up clock

Rivals are spending north of $100 billion each on AI infrastructure. Apple’s nine-month capex was $6.8 billion. If licensed intelligence is genuinely as good as owned intelligence, the Ternus era looks like the smartest capital allocation in the sector. If it is not, the gap compounds annually and cannot be closed by writing a bigger cheque to a supplier.

Data centres take years to build. That is the uncomfortable asymmetry of the Ternus era position: the decision to catch up cannot be executed in the quarter it becomes obvious.

Quarterly revenue, as reported: Apple to June 2026 against Nvidia to July 2026
Apple, total $109.4bn
Nvidia, total $96.2bn
Nvidia, data centre only $89.0bn
Apple, iPhone only $54.3bn

What the Ternus Era Means If You Are Buying AI

Neither of these companies is a template for a mid-sized organisation. Both are useful as a way of framing a decision most businesses now face.

The same question, at your scale

Every organisation adopting AI answers a version of this. Do you rent capability from a provider and concentrate your spending on the thing that makes you distinctive, or do you own more of the stack — models, infrastructure, tooling — and accept the cost and complexity that follow?

Apple picked the first at a scale where it could have afforded the second. That is worth noting before anyone assumes owning more is automatically the serious choice. Building an AI strategy starts with being honest about which layer actually differentiates you — the Ternus era is a very expensive demonstration that the answer is not always “all of them”.

What the Ternus era teaches about renting

Renting works when three things hold: the supplier is not your direct competitor in the same product, the cost scales predictably, and you retain a credible path to replacing them. Apple’s arrangement satisfies the third condition and is openly uncertain on the other two.

Apply that test to your own contracts. A model API from a vendor that also sells your product category, priced per token with usage you cannot forecast, is the same structure with fewer resources behind it — a Ternus era arrangement without Apple’s balance sheet to absorb the surprise.

What the Nvidia deal teaches about owning

Owning the stack works when the layers reinforce each other and you can absorb the regulatory attention that comes with concentration. Almost no organisation outside the largest handful can do both.

The practical read for everyone else is narrower. Nvidia buying the open model hub means the default place your engineers get models is now owned by your accelerator vendor. That is not a reason to change anything this quarter, but it is a reason to know which models you depend on and where else they can be sourced.

Three questions worth asking this quarter

Which parts of your AI stack would you have to rebuild if a supplier changed terms tomorrow? What does your inference bill do if usage triples? And is the capability you are renting the one your customers actually pay you for?

None of those requires a view on Apple or Nvidia. Both companies just spent a week demonstrating that the answers are not obvious even with unlimited money.

Ternus Era Questions People Are Asking

When did the Ternus era officially start?

The Ternus era started on 1 September 2026, when John Ternus became Apple’s chief executive and Tim Cook moved to executive chairman the same day. The transition had been announced in April 2026.

Is Apple building its own AI models?

Yes. Alongside the Google licence, Apple continues developing a cloud model of roughly a trillion parameters that could power Siri independently, reported as a possibility from next year. The Gemini arrangement is positioned as an interim step.

How much is Apple paying Google?

Reports put it at approximately $1 billion a year for a custom 1.2-trillion-parameter Gemini model. Apple has not published the terms.

Why is Nvidia buying software companies?

Because its thesis is that whoever controls where developers find and deploy models influences which accelerators get bought. Hugging Face at $12.93 billion is a distribution purchase, not a capability purchase.

Does the Hugging Face deal change open-source AI?

Not immediately. Nvidia committed to keeping the platform open across compute platforms and accelerators. The deal is not expected to close until the first half of 2027, and regulatory review runs until then.

What should businesses do differently?

Nothing urgent. Use the moment to document which models and platforms you depend on, what your inference costs do under growth, and how quickly you could switch. That inventory is useful regardless of how either bet turns out.

References